Language Language English 简体中文

Changelog / 变更日志

UniLab follows Semantic Versioning. This shared page records notable PyPI releases and unreleased changes in English and Chinese; release dates are PyPI upload dates. For the complete commit history, see the UniLab repository.

UniLab 遵循语义化版本。本共享页面以中英文记录重要的 PyPI 版本变更与未发布变更；发布日期采用 PyPI 上传日期。完整提交历史请参阅 UniLab 仓库。

1.5.1 (2026-10-23)

Added / 新增

• FlashSAC data-parallel learners now capture and replay the whole update cycle, including NCCL reductions, by default on supported compiled CUDA execution. Runner synchronizer hooks coordinate capture and replay, a packed loss-validity sentinel keeps optimizer and target updates rank consistent, and DP capture failure fails closed instead of allowing one rank to fall back to eager collectives. Set algo.algo_params.dp_whole_cycle_cuda_graph=false to disable DP capture, or algo.algo_params.use_whole_cycle_cuda_graph=false to disable capture on all ranks. training.dp_gradient_wire_dtype=bfloat16 additionally opts into lossy BF16 communication for sufficiently large FP32 gradient groups; FP32 remains the default and long-run convergence equivalence has not been established (#2499). FlashSAC 数据并行 learner 在受支持的编译 CUDA 执行路径上，现在默认捕获并回放 包含 NCCL reduction 的完整 update cycle。Runner synchronizer hook 协调捕获与 回放；打包的 loss-validity sentinel 保证 optimizer 与 target update 在各 rank 一致；DP 捕获失败会 fail closed，而不是允许单个 rank 回退到 eager collectives。 设置 algo.algo_params.dp_whole_cycle_cuda_graph=false 可关闭 DP 捕获； algo.algo_params.use_whole_cycle_cuda_graph=false 可在所有 rank 关闭捕获。 training.dp_gradient_wire_dtype=bfloat16 可额外为足够大的 FP32 gradient group 选择有损 BF16 通信；FP32 仍为默认，且尚未证明长时间训练收敛等价 （#2499）。

Changed / 变更

• Multi-GPU topology remains explicit and simple: omit CUDA_VISIBLE_DEVICES for one rank, or provide one entry per selected GPU. README now shows minimal two-GPU FlashSAC training and the manually managed CUDA MPS flow. An explicit training.cuda_process_sharing=mps request normalizes numeric GPU entries to host UUIDs before rank derivation and CUDA initialization, so the documented per-GPU MPS launch no longer fails on the second rank; uni-cumps env is reserved for external launchers (#2499, #2505). 多 GPU 拓扑保持显式且简单：省略 CUDA_VISIBLE_DEVICES 使用一个 rank；为每张 选择的 GPU 提供一个条目则启用对应 rank。README 现在展示最小双卡 FlashSAC 训练和手动管理的 CUDA MPS 流程。显式 training.cuda_process_sharing=mps 请求会在 rank 派生和 CUDA 初始化前把数字 GPU 条目规范化为主机 UUID，因此 文档中的 per-GPU MPS 启动不再在第二个 rank 失败；uni-cumps env 仅保留给 外部启动器使用（#2499、 #2505）。

Fixed / 修复

• PPO, APPO, SAC, and FlashSAC entrypoints now handle Ctrl+C cooperatively: runner and collector cleanup finishes, the first interrupt returns shell status 130 without a traceback, and a second interrupt forces termination. Actual worker and rank failures remain loud (#2503). PPO、APPO、SAC 与 FlashSAC 入口现在协作处理 Ctrl+C：runner 与 collector 清理完成后，第一次中断以 shell status 130 退出且不输出 traceback；第二次 中断强制终止。真实的 worker/rank 失败仍会完整暴露 （#2503）。

• The off-policy live terminal display now starts at the collector-ready handshake, preventing learner compilation, CUDA graph capture, and collector startup messages from interleaving with live redraws (#2507). off-policy live 终端显示现在延迟到 collector-ready handshake，避免 learner 编译、CUDA graph 捕获和 collector 启动消息与 live 刷新交错 （#2507）。

• Off-policy runners initialize the learner inference timer on every supported device and record the effective cuda_events or synchronized wall-clock implementation in the runtime manifest. This fixes macOS MPS SAC/FlashSAC training failing at the first inference request (#2509). off-policy runner 会在每个受支持的 learner 设备上初始化推理计时器，并把实际 使用的 cuda_events 或同步 wall-clock 实现记录到 runtime manifest。该修复解决 macOS MPS SAC/FlashSAC 训练在第一次推理请求时失败的问题 （#2509）。

• Fused community 13-D root-state and motion-state resets now rotate world-frame angular velocity into the generalized free-root/body frame before tensor submission, while linear velocity remains world-frame. Rotated roots with nonzero angular velocity therefore initialize the intended physical state without reintroducing host-boundary work (#2513). 融合 community 13-D root-state 与 motion-state reset 现在会在 tensor 提交前把 world-frame 角速度旋转到 generalized free-root/body frame，线速度仍保持 world-frame。带非零角速度的旋转 root 因此能初始化为预期物理状态，且没有重新 引入 host-boundary 开销（#2513）。

1.5.0 (2026-10-22)

Breaking changes / 破坏性变更

• An explicit training.collector_tensor_device=cpu paired with a CUDA learner now fails closed on non-ROCm Torch builds: the CPU carrier is no longer a CUDA-learner performance path (issue #2479 retest showed throughput parity with CUDA carriers plus CUDA MPS). CPU carriers remain legal for CPU and torch.mps learners and for ROCm/HIP builds; attribution baselines and debug runs opt in via UNILAB_ALLOW_CUDA_LEARNER_CPU_CARRIER=1 (#2483, ADR-0019). 显式 training.collector_tensor_device=cpu 与 CUDA learner 的组合在非 ROCm Torch 构建上现在 fail closed：CPU 载体不再是 CUDA learner 的性能路径 （#2479 重测确认其与 CUDA 载体 + CUDA MPS 吞吐持平）。CPU 载体对 CPU / torch.mps learner 与 ROCm/HIP 构建仍然合法；归因基线与调试 run 通过 UNILAB_ALLOW_CUDA_LEARNER_CPU_CARRIER=1 显式 opt-in（#2483，ADR-0019）。

• training.sim_carrier_device and training.inference_transport are removed without a compatibility layer; both merge into the single one-dimensional collector-tensor residency knob training.collector_tensor_device (auto/cpu/cuda, default auto, following the learner device). The off-policy inference transport is now a pure derived fact (cpu → host ring, cuda → CUDA ring). The env owner field is likewise renamed env.sim_carrier_device → env.collector_tensor_device. All removed keys fail closed with a migration hint; run_config.json records the mapping under device_resolution.collector_tensor (#2481). training.sim_carrier_device 与 training.inference_transport 已删除， 无兼容层；两者合并为单一一维 collector tensor 驻留旋钮 training.collector_tensor_device（auto/cpu/cuda，默认 auto， 跟随 learner 设备）。off-policy inference transport 变为纯派生事实 （cpu → host ring，cuda → CUDA ring）。env owner 字段同步改名 env.sim_carrier_device → env.collector_tensor_device。全部已移除键 带迁移提示 fail closed；run_config.json 在 device_resolution.collector_tensor 记录映射（#2481）。

• The public reset contract speaks IsaacLab/mjlab-style env_ids row indices again: TorchEnv.reset, ManagerBase.reset and all manager/term reset boundaries, EventManager.apply(mode="reset", env_ids=...), reset owners, and the recorder pre/post-reset hooks take a one-dimensional integer tensor (or None for every row). The v1.4 fixed-width bool mask form is removed without a compatibility layer: a bool tensor passed positionally now raises TypeError, mask-form reset event terms (reset(env, reset_mask, **params)) are no longer detected or dispatched, and the EventManager.dispatch_reset wrapper is gone — use apply(mode="reset", env_ids=...). The keyword form likewise raises TypeError; rename the argument to env_ids and pass row indices (#2460, #2474). 公共 reset 契约恢复 IsaacLab/mjlab 风格的 env_ids 行索引形态： TorchEnv.reset、ManagerBase.reset 及所有 manager/term reset 边界、 EventManager.apply(mode="reset", env_ids=...)、reset owner 与 recorder 的 pre/post-reset 钩子均接收一维整数张量（None 表示全部行）。v1.4 的 定宽 bool mask 形态已直接移除、无兼容层：位置传参的 bool 张量现在抛 TypeError；mask 形态的 reset 事件项（reset(env, reset_mask, **params)） 不再被检测或派发；EventManager.dispatch_reset 包装已删除——请使用 apply(mode="reset", env_ids=...)。关键字形式同样抛 TypeError；请将 参数名改为 env_ids 并传行索引（#2460，#2474）。 迁移步骤与逐类 term 的 before/after 示例见迁移指南 Migrating from UniLab 1.4 to 1.5 / 从 UniLab 1.4 迁移到 1.5。

Added / 新增

• training.tf32_matmul is a new PPO/APPO float32 matmul backend knob: ieee (default) preserves full-precision FP32, and tf32 opts into TF32 tensor-core kernels with an automatic IEEE fallback on MPS/macOS, ROCm/HIP, pre-Ampere GPUs (compute capability below 7.0), and CPU. Training entrypoints apply the policy before CUDA and learner construction, play paths honor it, and the resolved value plus fallback reason are recorded in run_config.json/run_summary.json. The knob is intentionally not exposed for SAC-family learners: their bf16-AMP updates already cast the main matmuls, and the closed #2477 benchmark measured TF32 as neutral there. The main benefit is the FP32 PPO/RSL-RL update path (#2475). 新增 PPO/APPO 的 float32 matmul 后端开关 training.tf32_matmul：ieee （默认）保持全精度 FP32；tf32 启用 TF32 tensor-core kernel，并在 MPS/macOS、ROCm/HIP、pre-Ampere GPU（compute capability < 7.0）与 CPU 上自动 回退为 IEEE。训练入口会在 CUDA 与 learner 构造前应用该策略，play 路径同样 生效，解析结果与回退原因写入 run_config.json/run_summary.json。该键有意 不暴露给 SAC 家族：其 bf16-AMP 更新已接管主要 matmul，已关闭的 #2477 实测 TF32 在该路径为噪声级。主要收益在 FP32 的 PPO/RSL-RL 更新路径（#2475）。

• Manager-owned observation history and delay are restored: ObservationTermCfg.history_length > 0 and delay_* construct and reset again, with the v1.3.4 semantics (oldest-to-newest flattening, first-frame and post-reset backfill, noise/clip/scale before history). Partial resets backfill the reset rows and peek without advancing other rows’ history, delay schedule, or lag RNG. CircularBuffer and DelayBuffer are now exported publicly from unilab.managers so downstream temporal owners no longer import the private unilab.managers._buffers package (#2462). 恢复 Manager 持有的观测 history/delay：ObservationTermCfg.history_length > 0 与 delay_* 重新可构造、可 reset，语义与 v1.3.4 一致（oldest→newest 展开、 首帧与 reset 后 backfill、noise/clip/scale 先于 history）。部分 reset 只对 reset 行做 backfill 并 peek，不推进其他行的 history、delay 日程或 lag RNG。 CircularBuffer 与 DelayBuffer 现从 unilab.managers 公开导出，下游 temporal owner 不再需要 import 私有包 unilab.managers._buffers（#2462）。

Changed / 变更

• The internal env_ids → bool mask adaptation layer now scatters into owner-held reusable buffers (unilab.base.reset_selectors.ResetMaskBuffer) instead of allocating a fresh num_envs bool tensor on every reset/event term call: the RL env reset entry, the event/observation/command managers, and per-term reset owners each hold their own scratch buffer, so the hot reset path no longer pays a per-call allocation. Public contracts are unchanged; reset_env_ids_mask keeps its allocating form for fixtures and external callers (#2461). 内部 env_ids → bool mask 适配层改为 scatter 进 owner 持有的可复用 buffer（unilab.base.reset_selectors.ResetMaskBuffer），不再在每次 reset/event term 调用时新分配 num_envs 大小的 bool 张量：RL env reset 入口、event/observation/command manager 及各 term 的 reset owner 分别持有 各自的 scratch buffer，热路径上消除了逐次分配。公开契约不变； reset_env_ids_mask 保留逐次分配形态供 fixture 与外部调用方使用（#2461）。

Fixed / 修复

• Fixed the mjwarp off-policy collector Env Step breakdown hiding Backend Step: an event-backed asynchronous step’s physics wait surfaces at the first host read of the step outputs — the update-state termination drain, or the autoreset done-mask gate when update_state performs no host read — and was misattributed to Update State / Reset Done. Both windows now roll up into Backend Step through TorchEnv._collect_step_core_drain_ms; total Env Step accounting is unchanged and synchronous backends keep their existing attribution. The off-policy terminal System column also gained a CUDA MPS indicator (on / ambient / off) driven by the run’s cuda_process_sharing manifest evidence (#2495). 修复 mjwarp off-policy collector Env Step 分解缺 Backend Step 的问题： 事件型异步 step 的物理等待发生在 step 输出后的第一个 host 读取处 （update_state 内 termination drain；若 update_state 无 host 读取则为 autoreset done-mask 门控），此前被误归入 Update State / Reset Done。 两个窗口现在经 TorchEnv._collect_step_core_drain_ms 归入 Backend Step； Env Step 总额不变，同步后端归因不变。off-policy 终端 System 列新增 CUDA MPS 指示（on / ambient / off），取值自 run 的 cuda_process_sharing manifest 证据（#2495）。

1.4.0 (2026-10-20)

Breaking changes / 破坏性变更

• Resuming PPO/APPO training now fails closed on pre-patch rsl-rl checkpoints: a checkpoint without a valid unilab_logger_state — and, for environments that publish training state, without a valid versioned unilab_env_training_state — is rejected with an error naming the minimum supported format instead of silently restarting progress from 0. There is no compatibility window, offline migration tool, or dual-write path; retrain or resume from a checkpoint saved by a supported UniLab version (#2416). PPO/APPO 续训对 pre-patch rsl-rl checkpoint fail closed：缺少有效 unilab_logger_state 的 checkpoint——对于发布 training state 的环境，还需有效的 版本化 unilab_env_training_state——会被拒绝加载，错误信息说明最低支持格式， 不再静默地从 0 恢复进度。无兼容窗口、离线迁移工具或 dual-write 路径；请用受支持 的 UniLab 版本重训，或从该版本保存的 checkpoint 续训（#2416）。

• The sim2sim contract guard no longer warning-skips historical runs: a source run whose run_config.json is missing, malformed, lacks a mapping contract_snapshot, or contains keys outside the current contract schema fails closed with CrossBackendIncompatibleError. The legacy snapshot/config key aliases (env.sampling_mode, algo.policy.actor_hidden_dims, algo.policy.critic_hidden_dims, algo.empirical_normalization) are removed; only canonical contract_snapshot mappings written by training runs on the 1.3.5 baseline schema are supported, and this change is not backported to 1.3.5 patch releases. To reuse an older run, retrain with the current version (checkpoint weights are unchanged); training.sim2sim_strict=false does not bypass the integrity failures (#2416). Sim2Sim 契约守卫不再对历史 run warning-skip：源 run 的 run_config.json 缺失、损坏、缺少 mapping 形态的 contract_snapshot、或含有当前契约 schema 之外的键时一律 fail closed（抛 CrossBackendIncompatibleError）。legacy 快照/配置键别名（env.sampling_mode、algo.policy.actor_hidden_dims、 algo.policy.critic_hidden_dims、algo.empirical_normalization）已移除； 仅支持 1.3.5 基线 schema 训练 run 写出的 canonical contract_snapshot， 且本变更不回 port 到 1.3.5 patch release。要复用旧 run 需用当前版本重训 （checkpoint 权重本身不变）；training.sim2sim_strict=false 不能绕过完整性 失败（#2416）。

• FlashSAC pre-contract checkpoints (a missing or null actor_semantics entry) are no longer supported: training resume and play loading fail closed for every actor, including the built-in actor, with an error that states the minimum supported checkpoint format. The only migration path is to retrain the policy with a current release; the change is not backported to any 1.3.5 patch release (#2416). FlashSAC pre-contract checkpoint（缺失或为 null 的 actor_semantics 条目） 不再受支持：训练续训与 play 加载对所有 actor（包括内置 actor）一律 fail closed，错误信息说明最低支持的 checkpoint 格式。唯一迁移路径是使用当前 版本重新训练策略；该变更不回 port 到任何 1.3.5 patch release（#2416）。

• The training seed contract no longer falls back to training.seed; only the algorithm-level algo.seed is honored. A config that sets training.seed without algo.seed fails closed with an error pointing to algo.seed (#2415). 训练 seed 契约不再 fallback 到 training.seed，仅认可算法级 algo.seed。配置了 training.seed 但缺少 algo.seed 的配置会 fail closed，错误信息指向 algo.seed（#2415）。

• render_play_mode() no longer accepts the ignored sim_backend keyword; backend routing belongs to env.run_playback(). Callers that still pass the argument now get a TypeError; deleting the argument migrates the call with unchanged behavior (#2415). render_play_mode() 不再接受无效的 sim_backend 关键字参数；backend 路由 归属 env.run_playback()。仍传该参数的调用方会得到 TypeError；删除该 实参即可完成迁移，行为不变（#2415）。

• Interactive playback no longer supports the non-Hydra cfg=None path: create_playback_session() requires a composed Hydra DictConfig, and the sim2sim preflight fails closed instead of being silently skipped when the config is missing or is not a DictConfig (#2415). 交互式回放不再支持非 Hydra 的 cfg=None 路径：create_playback_session() 要求组合后的 Hydra DictConfig；config 缺失或不是 DictConfig 时 sim2sim preflight fail closed，不再静默跳过（#2415）。

• The Isaac setup scripts accept only UNISIM_ISAACGYM_HOME / UNISIM_ISAACSIM_HOME; the retired UNILAB_ISAACGYM_HOME / UNILAB_ISAACSIM_HOME spellings fail closed with an error naming the current variables (#2415). Isaac 安装脚本仅接受 UNISIM_ISAACGYM_HOME / UNISIM_ISAACSIM_HOME；已 退役的 UNILAB_ISAACGYM_HOME / UNILAB_ISAACSIM_HOME 拼写会 fail closed，错误信息指向当前变量名（#2415）。

• unilab-pull-assets no longer accepts the no-op --verbose option; passing it now fails argument parsing. Output stays compact by default (#2415). unilab-pull-assets 删除 no-op 的 --verbose 选项；传入该选项会导致参数 解析失败。输出默认保持紧凑（#2415）。

• Benchmark tooling drops its legacy compatibility surface: physics benchmarks accept only the canonical owner task ids (go2_joystick_flat, g1_walk_flat; env-class names, route forms, and short names are rejected), replay placement accepts only --tasks (the singular --task alias is removed), the SAC replay-sampling benchmark’s legacy defaults and positional sim argument are removed, the IsaacGym/IsaacSim benchmark scripts drop the UNILAB_ISAACGYM_DEPS_ROOT / UNILAB_ISAACGYM_MODELS_ROOT environment aliases and machine-local path defaults (unset paths fail closed), and the Torch benchmark workloads publish a single final_observation carrier instead of the dual-written compat_final_obs buffer (#2413). benchmark 工具链删除 legacy 兼容面：physics benchmark 仅接受 canonical owner task id（go2_joystick_flat、g1_walk_flat；env-class 名、路由形式 与短名一律拒绝）；replay placement 仅接受 --tasks（删除单数 --task 别名）；SAC replay sampling benchmark 删除旧默认值与位置参数 sim 解释； IsaacGym/IsaacSim benchmark 脚本删除 UNILAB_ISAACGYM_DEPS_ROOT / UNILAB_ISAACGYM_MODELS_ROOT 环境变量别名与机器本地路径默认值（未设置时 fail closed）；Torch benchmark 工作负载只发布单一 final_observation 载体，不再双写 compat_final_obs 缓冲（#2413）。

• training.collector_tensor_device, manager_torch_device, env.tensor_runtime, and env.tensor_runtime_device are removed. Use training.sim_carrier_device (routed to env.sim_carrier_device) and the resolved learner device for tensor-native inference. Removed keys fail closed with migration hints, cuda no longer maps silently to MPS, and run_config.json records configured-to-resolved devices. The required unisim-core 1.7.25 enum support inventory and runtime parity guard are pinned (#2382). training.collector_tensor_device、manager_torch_device、env.tensor_runtime 与 env.tensor_runtime_device 已移除。请使用 training.sim_carrier_device（路由到 env.sim_carrier_device），并让 tensor-native 推理跟随已解析 learner 设备。移除键会带迁移提示 fail closed，cuda 不再静默映射 MPS，run_config.json 会记录“配置值 → 解析值”设备映射。所需 unisim-core 1.7.25 枚举支持清单与 runtime parity 守卫已固定。

• The former unilab-rl distribution (import namespace uni_rl) is merged into the unilab package under unilab.rl (contracts / adapters / algos / runtime / logging layers), and the release surface converges to the single unilab distribution. The transitional uni_rl extra has been removed: legacy --extra uni_rl / pip install unilab[uni_rl] commands are no longer valid, and permanent import uni_rl support is not guaranteed. See ADR-0016 RL Reintegration Into unilab.rl And Single Distribution and the uni_rl.* → unilab.rl.* import-path mapping in Migrating from unilab-rl (uni_rl) (#2372). 原 unilab-rl distribution（import namespace uni_rl）并入 unilab 包的 unilab.rl（contracts / adapters / algos / runtime / logging 分层）， 发行面收敛为单一 unilab distribution。过渡期 uni_rl extra 已移除， 旧 --extra uni_rl / pip install unilab[uni_rl] 命令不再有效，且不承诺 永久支持 import uni_rl。决策见 ADR-0016，uni_rl.* → unilab.rl.* 的 import 路径映射见迁移页 从 unilab-rl（uni_rl）迁移（#2372）。

• Removed the collector CUDA-graph public lifecycle and graph-named completion contract without aliases or compatibility paths. Runtime code now consumes the neutral completion-v2 receipt (state.info["completion"]) while eager ordered visibility, producer validation, actual-operation waits, backend-owned physics graphs, and learner CUDA graphs remain in their owner layers. Each successful step publishes its current receipt, and an autoreset step publishes the reset boundary receipt that gates its post-reset observation carrier (#2359, #2358). 移除 collector CUDA graph 的公共 lifecycle 与 graph 命名 completion 契约， 不保留 alias 或兼容路径。运行时改用中性 completion-v2 receipt （state.info["completion"]），同时保留 eager 有序可见性、producer 校验、 按实际操作等待、backend 私有 physics graph 与 learner CUDA graph （#2359、#2358）。

Changed / 变更

• Raised the required unisim-core pin to 1.8.2 for the base, Motrix, and SuperDex dependency profiles, and removed the issue-local editable UniSim source override. The published release supplies the restored packed-reset, Isaac semantic-boundary, and Motrix wxyz FrameQuat contracts. 基础、Motrix 与 SuperDex 依赖配置将 unisim-core 提升到 1.8.2，并移除 issue 本地的 editable UniSim source override。已发布版本提供恢复后的 packed-reset、Isaac 语义边界与 Motrix wxyz FrameQuat 契约。

• training.collector_tensor_device now defaults to auto: MuJoCo HOST_BRIDGE training follows the learner device, so a CUDA learner receives CUDA Manager/TorchEnv carriers and CUDA inference transport without explicit overrides, while CPU and torch.mps (Metal Performance Shaders) learners keep the CPU carrier. Explicit cpu/cuda values are unchanged and win over auto. Existing MuJoCo configurations that relied on the previous cpu default with a CUDA learner now allocate one CUDA context per collector process on the learner GPU; pin training.collector_tensor_device=cpu to keep the previous topology (#2365). training.collector_tensor_device 默认值改为 auto：MuJoCo HOST_BRIDGE 训练跟随 learner 设备——CUDA learner 无需显式配置即可获得 CUDA Manager/TorchEnv carrier 与 CUDA inference transport，CPU 与 torch.mps （Metal Performance Shaders）learner 保持 CPU carrier。显式 cpu/cuda 语义不变且优先于 auto。原先在 CUDA learner 下依赖 cpu 默认值的 MuJoCo 配置，现在每个 collector 进程会在 learner GPU 上额外占用一个 CUDA context；如需保持原拓扑，请显式固定 training.collector_tensor_device=cpu（#2365）。

Fixed / 修复

• Fixed capability-scoped Manager carrier routing for every in-process packed HOST_BRIDGE backend, not only MuJoCo. Explicit training.sim_carrier_device=cuda and the default auto routing now honor the public backend torch_devices declaration for Motrix, Drake, and SuperDex, while unsupported or unknown device declarations fail closed. Physics remains CPU-authoritative; this changes only public Manager carrier placement (#2401). 修复 Manager carrier 路由仅限 MuJoCo 的问题：显式 training.sim_carrier_device=cuda 与默认 auto 现在按公开的后端 torch_devices 声明作用于 Motrix、Drake 与 SuperDex 等全部 in-process packed HOST_BRIDGE 后端；未声明或不受支持的设备会 fail closed。物理仍 以 CPU 为权威，本修复只改变公共 Manager carrier 放置位置。

• Fixed device-resident Manager sensor staleness and restored SAC g1_walk_flat policy quality across the newly adapted backends. Every compiled derived sensor now crosses its backend boundary on full and selected-reset reads; IsaacSim uses a task-wide finite floor with audited 50 m clone spacing; and IsaacGym clamps position targets to source MJCF joint bounds. Full 5000-iteration SAC runs recover Genesis/Newton/IsaacSim/IsaacGym rewards and episode lengths comparable with the MuJoCo-family baselines (#2387). 修复 device-resident Manager sensor 停滞问题，并恢复新适配后端在 SAC g1_walk_flat 上的策略质量。每个 compiled derived sensor 在 full 与 selected-reset 读取时都会重新跨越 backend 边界；IsaacSim 使用覆盖任务的 finite floor 并审计 50 m clone 间距；IsaacGym 将 position target clamp 到 source MJCF joint bound。四条 5000-iteration SAC 完整运行的 reward 与 episode length 已回到 MuJoCo-family baseline 区间（#2387）。

• Off-policy replay timeout bootstrapping: a transition that returns terminated=True and truncated=True together is a true terminal and no longer bootstraps. The collector now writes the timeout-bootstrap eligibility mask (truncated & ~terminated) into the replay truncated column while dones keeps the terminated | truncated lifecycle semantics and done rows keep the real terminal observations; SAC/FlashSAC targets are unchanged in form and now correct for the overlap case (unilab_rl#29). 修复 off-policy replay 的 timeout bootstrap：同一 transition 同时返回 terminated=True 与 truncated=True 时是真正的终止，不再 bootstrap。 collector 现在把 timeout bootstrap 资格掩码（truncated & ~terminated） 写入 replay 的 truncated 列；dones 保持 terminated | truncated 的生命周期语义，done 行仍携带真实 terminal observation；SAC/FlashSAC 的 target 公式形式不变，overlap 情形结果修正（unilab_rl#29）。

• Multi-term concatenated observations are again caller-owned: the Manager no longer publishes rotating internal output carriers through the default eager boundary, so observations retained by a transition consumer remain immutable across later steps (#2362). 多 term 拼接 observation 重新归 caller 所有：Manager 不再通过默认 eager boundary 发布轮换的内部 output carrier，因此 transition consumer 保留的 observation 不会在后续 step 中被改写（#2362）。

• Restored the v1.3.5 MotionCommand empty-reset-mask semantics: an empty mask consumes no sampler RNG and performs no backend state write or packet ingestion (#2359). 恢复 v1.3.5 的 MotionCommand 空 reset mask 语义：空 mask 不消耗 sampler RNG，也不写 backend state 或摄取 packet（#2359）。

• Fixed IsaacGym/IsaacSim interactive eval, which failed before the native viewer opened: unisim-core 2.0.0 fails closed when an Isaac native interactive viewer receives a non-default camera_kwargs profile, while the eval play path forwards training.cam_* to every backend. run_playback_mode now mirrors the interactive debug-overlay degradation — it warns, drops the camera profile, and retries interactive playback; record mode still forwards the profile and fails closed on rejection (#2454). 修复 IsaacGym/IsaacSim 交互 eval 在原生 viewer 打开前失败的问题： unisim-core 2.0.0 对收到非默认 camera_kwargs 的 Isaac 原生交互 viewer fail closed，而 eval 播放路径会把 training.cam_* 传给所有后端。 run_playback_mode 现在与 interactive debug-overlay 的降级模式一致—— 告警、去掉相机配置并重试交互播放；record 模式仍照常下发相机配置， 被拒绝时 fail closed（#2454）。

1.3.5 (2026-10-13)

Breaking changes / 破坏性变更

• Completed the production Torch-runtime closeout: UniLab no longer imports, compiles, tunes, or requires Numba. The CPU motion command and kernels were removed, cpu_ids owns only Linux process affinity, and every G1 motion owner uses the device-resident MotionCommandCfg. Manager, env, and task runtime code uses Torch-authoritative carriers and rejects hidden Torch↔NumPy conversions; the cold NPZ MotionLoader is the only task-runtime NumPy leaf, while base owners confine legacy NumPy exchange to explicit adapter/reset-host boundaries (#2124, ADR-0015). Production Numba 实现与依赖已完整移除：UniLab 不再 import、编译、调优或 依赖 Numba。CPU motion command 与 kernels 已删除，cpu_ids 仅保留 Linux process affinity 语义，所有 G1 motion owner 均使用 device-resident MotionCommandCfg。Manager、env 与 task runtime 使用 Torch-authoritative carrier，并拒绝隐藏的 Torch↔NumPy 转换；cold NPZ MotionLoader 是 task runtime 唯一 NumPy leaf，base owner 只在显式 adapter/reset-host boundary 内消化 legacy NumPy 交换（#2124、ADR-0015）。

• The Manager-Based reset lifecycle now fails closed unless the backend declares packed_reset. UniLab’s NumPy-authoritative reset composer and non-packed fallback were removed: reset-mode state, model-field, and mocap writes are composed through UniSim’s public packed transaction and committed once per lifecycle. UniLab no longer calls SimBackend.set_state_tensor_mask(); mapped-entity tensor commits remain an explicit migration boundary (#2275, #2280, #2289). Manager-Based reset lifecycle 现在要求 backend 声明 packed_reset，否则 fail closed。UniLab 侧 NumPy-authoritative reset composer 与 non-packed fallback 已删除：reset-mode 状态、模型字段与 mocap 写入通过 UniSim 公共 packed transaction 组合，并在每个生命周期恰好提交一次。UniLab 不再调用 SimBackend.set_state_tensor_mask()；mapped-entity tensor 提交仍是显式 迁移边界（#2275、#2280、#2289）。

• The training seed contract moved from the unilab.utils.seed transition shim to its owner at unilab.training.seed, without a compatibility alias. unilab.training continues to re-export the owner-facing helpers. algo.seed now seeds only global Python and Torch RNGs, and random interactive playback samples actions with Torch (#2124, #2283). 训练 seed 契约已从 unilab.utils.seed 过渡 shim 迁移到 owner unilab.training.seed，且不保留兼容 alias。unilab.training 继续导出 owner-facing helper。algo.seed 现在只 seed 全局 Python 与 Torch RNG， 随机 interactive playback 改用 Torch 采样 action（#2124、#2283）。

• Removed the unilab.utils.rotation transition shim without a compatibility alias. Base-owner host math moved to unilab.base.rotation, while motion NPZ continuity and angular-velocity conversion now live inside the cold MotionLoader owner. 过渡 shim unilab.utils.rotation 已移除且不保留兼容 alias。base-owner host 数学迁移到 unilab.base.rotation；motion NPZ 连续性与角速度转换则 收进 cold MotionLoader owner 内部。

Added / 新增

• Added backend tensor-runtime diagnostics. TorchEnv serializes the public backend diagnostics, training records them in runtime_manifest.backend_diagnostics, and direct PPO runs publish the same payload as top-level backend_diagnostics in run_summary.json; unavailable backends and probe failures are recorded explicitly (#2278, #2298). 新增 backend tensor-runtime diagnostics。TorchEnv 会序列化公共 backend diagnostics，训练将其记录到 runtime_manifest.backend_diagnostics，direct PPO run 会在 run_summary.json 顶层发布同一 payload；backend 不支持或探测失败都会显式 记录（#2278、#2298）。

Changed / 变更

• Manager terms and tasks consume semantic, packed tensor state views for joint, body, root, contact, height, and sensor data. Legacy NumPy backend methods are no longer UniLab runtime APIs, and Entity host exchange is confined to the backend adapter boundary (#2276, #2277). Manager term 与 task 改为消费 joint、body、root、contact、height 与 sensor 数据的 semantic packed tensor view。legacy NumPy backend 方法不再是 UniLab runtime API，Entity host 交换收进 backend adapter 边界 （#2276、#2277）。

• Selected reset execution replaced dynamically materialized row carriers with fixed-width packed masks. Commands, events, metrics, observations, terminations, curricula, and task state implement explicit mask-reset boundaries, commit only touched rows, and preserve active state elsewhere (#2285, #2288). selected reset 执行以 fixed-width packed mask 取代动态物化 row carrier。 command、event、metric、observation、termination、curriculum 与 task state 均实现显式 mask-reset 边界，只提交被写入的行，并保留其他行的 active 状态 （#2285、#2288）。

• UniSim now owns packed reset composition: selected-row device qpos/qvel staging, selected-column merges, committed model-field baselines, startup baseline promotion, selected mocap staging, and the single public commit boundary. The composer is cached per declared device and its baseline persists across resets (#2275). UniSim 现在拥有 packed reset composition：selected-row device qpos/qvel staging、选中列合并、已提交模型字段 baseline、startup baseline 提升、 selected mocap staging，以及唯一公共提交边界。composer 按声明 device 缓存，baseline 跨 reset 持久（#2275）。

• Required runtime dependencies moved to unisim-core>=1.7.17 and unilab-rl==1.4.12; the default and ROCm dependency profiles are aligned (#2281, #2299, #2301). 必需运行时依赖升级为 unisim-core>=1.7.17 与 unilab-rl==1.4.12； default 与 ROCm dependency profile 已对齐（#2281、#2299、#2301）。

Fixed / 修复

• Fixed MuJoCo reset model-field DR commits through the host bridge, including startup randomization baselines (#2122, #2123). 修复 MuJoCo reset 模型字段 DR 经 host bridge 的提交，包括 startup randomization baseline（#2122、#2123）。

• Fixed canonical off-policy packed-reset paths: root and motion payloads are normalized to contiguous float32 tensors, owner-subset commands track their actual fixed-width write mask while preserving unwritten active rows, and SAC MuJoCo PD gains compose as final kp/kd model fields instead of silently skipping randomization (#2296, #2300). 修复典型 off-policy packed-reset 路径：root 与 motion payload 归一化为 contiguous float32 tensor，owner-subset command 记录实际 fixed-width 写入 mask 并保留未写入的 active rows，SAC MuJoCo PD gain 以最终 kp/kd 模型字段 组合，而不是静默跳过随机化（#2296、#2300）。

• Fixed motion NPZ body-column selection by local IDs, CUDA Manager carrier placement, and G1 reset-row observation boundaries (#2124, #2279, #2303). 修复按 local ID 选择 motion NPZ body column、CUDA Manager carrier placement， 以及 G1 reset-row observation 边界（#2124、#2279、#2303）。

1.3.4 (2026-10-10)

Breaking changes / 破坏性变更

• The Manager-Based runtime is now tensor-only: NpEnv and NpEnvState were removed, TorchEnv is the sole environment lifecycle base, and Manager state, observations, rewards, terminations, commands, events, metrics, and recorder carriers use Torch tensors. Hot-path step/reset inputs no longer accept NumPy arrays or perform hidden host conversion (ADR-0011, ADR-0012). Manager-Based 运行时改为 tensor-only：删除 NpEnv / NpEnvState， TorchEnv 成为唯一环境生命周期基类；Manager 的状态、观测、奖励、终止、 command、event、metric 与 recorder carrier 均使用 Torch tensor。热路径 step/reset 不再接受 NumPy 输入，也不再做隐藏 host 转换（ADR-0011、 ADR-0012）。

• Removed the production Numba implementation and dependency: the CPU motion command and its kernels were removed, cpu_ids now owns only Linux process affinity, and all G1 motion owners use the device-resident MotionCommandCfg. Manager/env runtime packages reject NumPy imports and hidden Torch↔NumPy conversions; the cold NPZ MotionLoader remains the only task-runtime NumPy leaf, while base owners consume legacy NumPy formats only in explicit adapter/reset-host boundaries (#2124, ADR-0015). 移除 production Numba 实现与依赖：删除 CPU motion command 和 kernels， cpu_ids 仅保留 Linux process affinity 语义，所有 G1 motion owner 均使用 device-resident MotionCommandCfg。Manager/env runtime package 禁止 NumPy import 与隐藏 Torch↔NumPy 转换；cold NPZ MotionLoader 仍是 task runtime 唯一 NumPy leaf，base owner 只在显式 adapter/reset-host boundary 内消化 legacy NumPy format（#2124、ADR-0015）。

• The training-entry seed contract no longer mutates the legacy global NumPy RNG; algo.seed seeds global Python and Torch RNGs only. Random interactive playback also samples its actions with Torch instead of np.random.uniform. This removes the last environment-owned global NumPy RNG semantics from training (#2124). 训练入口 seed 契约不再修改 legacy 全局 NumPy RNG；algo.seed 只 seed 全局 Python 与 Torch RNG。随机 interactive playback 也改用 Torch 采样 action，不再调用 np.random.uniform。该变更移除训练中最后一个 environment-owned 全局 NumPy RNG 语义（#2124）。

• The transition-shim module unilab.utils.rotation was removed without a compatibility alias. Its base-owner host math moved to unilab.base.rotation, while motion NPZ continuity and angular-velocity conversion now live inside the cold MotionLoader owner. 过渡 shim unilab.utils.rotation 已移除且不保留兼容 alias。base-owner host 数学迁移到 unilab.base.rotation；motion NPZ 连续性与角速度转换则 收进 cold MotionLoader owner 内部。

• Production training now enters through the sole Manager-Based factory and the scoped tensor backend set. Retired task owners, the WarpSAC config tree, and task-owned direct training environment classes were removed; fused task components remain Manager-owned terms rather than alternate environments (#2071, #2072, ADR-0012). 生产训练统一进入唯一的 Manager-Based factory 与 scoped tensor backend 集。 退役 task owner、WarpSAC 配置树和 task-owned direct environment class 均已 移除；融合任务组件保持为 Manager-owned term，而不是另一套环境类。

• Required runtime dependencies moved to unisim-core>=1.7.15, unilab-rl==1.4.11, and mjbatch-uni~=0.2.5; SAC uses the upstream SAC learner naming (algo_log_name=sac) (#2071, #2074, #2112, #2115). 必需运行时依赖升级为 unisim-core>=1.7.15、unilab-rl==1.4.11、 mjbatch-uni~=0.2.5；SAC 使用上游 SAC learner 命名 （algo_log_name=sac）。

Added / 新增

• Added a Manager-owned selected-reset lifecycle: tensor reset rows, an owner reset transaction, reset-row observation publication, device-resident reset randomization negotiated through device_reset_randomization, and startup-mode body mass/COM randomization with committed baseline promotion. 新增 Manager-owned selected-reset 生命周期：tensor reset row、owner reset transaction、reset-row observation publication、通过 device_reset_randomization 协商的 device-resident reset DR，以及带已提交 baseline promotion 的 startup-mode body mass/COM DR。

• Added scene-owned packed tensor reads for entity joint, body, and named sensor state; Manager terms consume validated device views, stable aggregate views, and backend-declared DEVICE_RESIDENT sensor inventories instead of hardcoded backend tables. Missing required device sensors fail closed. 新增 scene-owned packed tensor read：覆盖 entity joint/body/named sensor 状态；Manager term 消费经过校验的 device view、稳定 aggregate view 与后端 声明的 DEVICE_RESIDENT sensor inventory，而不是硬编码后端表。缺失的必需 device sensor 会 fail closed。

• Added Manager-owned Torch RNG and deterministic seed/reset behavior for observation noise, commands, events, impulses, and selected reset. Device noise, temporal/delay state, and scalar sensor observation carriers are tensor-native. 新增 Manager-owned Torch RNG，覆盖观测噪声、command、event、impulse 与 selected reset 的 seed/reset 行为。噪声、temporal/delay 状态和标量 sensor observation carrier 均为 tensor-native。

• Added the CUDA MPS single-rank execution-sharing mode and the user-owned uni-cumps CLI (status, doctor, start, stop, env). Training requests sharing explicitly with training.cuda_process_sharing=mps and validates a recorded daemon before construction; it never silently starts or stops a host service (ADR-0013, ADR-0014, #2065, #2069, #2085). 新增 CUDA MPS single-rank execution-sharing 模式与用户拥有的 uni-cumps CLI（status / doctor / start / stop / env）。训练必须通过 training.cuda_process_sharing=mps 显式请求，并在构造前校验已记录 daemon； 训练进程不会隐式启动或停止 host 服务（ADR-0013、ADR-0014）。

• Added tensor-runtime production soak tooling, replay-ingress diagnostics, abnormal-shutdown evidence, schema/bounds consumption, tensor NaN guards, and the production guide for the scoped G1 Motion Tracking / FlashSAC / MJWarp path. 新增 tensor-runtime 生产 soak 工具、replay-ingress 诊断、异常退出证据、 schema/bounds 消费、tensor NaN guard，以及 scoped G1 Motion Tracking / FlashSAC / MJWarp 路径的生产指南。

• Added opt-in off-policy training resume (algo.resume=true), explicit playback checkpoint selection, FlashSAC policy_before_critic, learner whole-cycle CUDA Graph controls, and grouped terminal metrics with bounded width tiers. 新增 opt-in off-policy 训练续训（algo.resume=true）、显式 playback checkpoint 选择、FlashSAC policy_before_critic、learner whole-cycle CUDA Graph 控制，以及带宽度分层的分组终端指标。

• Added generic motion-tracking capabilities retained for downstream task packages: universal combo rewards, optional motion NPZ torque fields, tensor motion sampler/carriers, initialized physical action targets with public tensor_target, Gaussian-noise clamp, recursive hashable keys for shared cross-group observation terms, and MJWarp set_state_tensor benchmark timings/profile syntax. 新增供下游任务包复用的 generic motion-tracking 能力：universal combo reward、可选 motion NPZ torque 字段、tensor motion sampler/carrier、确定性 初始化且公开 tensor_target 的物理 action target、Gaussian noise clamp、 跨组共享 observation term 的递归 hashable key，以及 MJWarp set_state_tensor benchmark timing/profile 语法。

Changed / 变更

• Reduced Manager host boundaries through packed reads, stable device views, fused action/reward/termination/observation/reset kernels, device-side samplers and RNG, cached tensor bounds, delayed reward host publication, and phase-level update-state/reset timing attribution. 通过 packed read、稳定 device view、融合 action/reward/termination/ observation/reset kernel、device-side sampler 与 RNG、缓存 tensor bounds、 延迟 reward host publication，以及 update-state/reset 的阶段级 timing attribution，减少 Manager host boundary。

• MuJoCo remains the canonical HOST_BRIDGE backend. Its default Manager carrier stays on CPU even when a learner resolves to CUDA; an indexed CUDA collector carrier remains an explicit training.collector_tensor_device=cuda request. MuJoCo 仍是 canonical HOST_BRIDGE backend。即使 learner 解析到 CUDA，其默认 Manager carrier 也保持 CPU；indexed CUDA collector carrier 必须显式设置 training.collector_tensor_device=cuda。

• Removed the repository-wide uv version pin and kept lockfile reproducibility; CI and Docker use the standard uv installation path. 移除仓库级 uv 版本限制；依赖可复现性仍由 lockfile 提供，CI 与 Docker 使用 标准 uv 安装路径。

Fixed / 修复

• Moved MimicLite/FlashMimic task ownership downstream: UniLab no longer ships the three mjwarp_mimiclite* owners, task torque-sensor scene, future-window observation module, or MimicLite applied-torque term. Generic tensor runtime capabilities remain upstream (#2119, #2120). MimicLite/FlashMimic 任务所有权迁至下游：UniLab 不再内置三个 mjwarp_mimiclite* owner、任务 torque-sensor 场景、future-window observation module 或 MimicLite applied-torque term；generic tensor runtime 能力保留在上游。

• Fixed frozen full-refresh motion references, partial tensor-command probing, stale device-resident joint reward state, selected-reset readiness boundaries, initial episode counter staggering, reset-row observation scale slicing, and device sensor fallback behavior. 修复 full refresh 中 motion reference 冻结、partial tensor-command probing、 joint reward 使用 stale device-resident state、selected-reset readiness boundary、初始 episode counter 交错、reset-row observation scale 切片，以及 device sensor fallback 行为。

• Fixed host-bridge Manager Torch device handling, explicit device validation, APPO/MJWarp test stability, ONNX export device examples, IsaacSim mapped owner cold/dtype/readiness contracts, and Motrix owner restoration. 修复 host-bridge Manager Torch device 处理、显式设备校验、APPO/MJWarp 测试 稳定性、ONNX export device 示例、IsaacSim mapped owner 的 cold/dtype/ readiness 契约，以及 Motrix owner 恢复。

1.3.3 (2026-09-27)

Changed / 变更

• Removed the repository-wide uv version pin: projects no longer require uv==0.12.5, CI and Docker use the standard uv installation path, and dependency reproducibility remains provided by uv.lock (#1809). 移除仓库级 uv 版本限制：项目不再要求 uv==0.12.5，CI 与 Docker 使用 标准 uv 安装路径；依赖可复现性仍由 uv.lock 提供（#1809）。

Added / 新增

• Added the WarpSAC off-policy algorithm shipped by unilab-rl 1.4.0: the conf/warpsac owner config tree with G1 owners for g1_walk_flat and g1_motion_tracking on the MuJoCo and mjwarp backends, the train_warpsac.py entrypoint, CLI and documentation wiring, and a stabilized G1 walk replay profile; explicit WarpSAC playback run selection is now honored (#1645). 新增 unilab-rl 1.4.0 提供的 WarpSAC off-policy 算法：conf/warpsac owner 配置树，覆盖 MuJoCo 与 mjwarp 后端的 g1_walk_flat 与 g1_motion_tracking G1 owner、train_warpsac.py 入口、CLI 与文档接线， 以及稳定化的 G1 walk 回放 profile；显式指定的 WarpSAC 回放 run 现在会被 正确遵循（#1645）。

• The viser browser-based viewer is now a required dependency and works on every playback-capable backend: eval --render-mode viser and training.play_render_mode=viser route through the shared playback loop for mujoco/mjwarp/newton/drake/superdex (#1631, #1632, #1638), and the Motrix backend gained viser rendering through the unisim-core 1.7.6 physics-state playback contract (portable entity scenes and audited whole-MJCF scenes) plus the 1.7.8 <worldbody> merge fix (#1639). The viser.* config group (port, max_envs, display_mode, env_idx) is now available on the appo/sac/flashsac configs to match ppo. viser 浏览器查看器成为必需依赖，并支持所有具备回放能力的后端： eval --render-mode viser 与 training.play_render_mode=viser 通过共享 回放循环支持 mujoco/mjwarp/newton/drake/superdex（#1631、#1632、#1638）； Motrix 后端借助 unisim-core 1.7.6 的 physics-state 回放契约（portable 实体场景与经审计的 whole-MJCF 场景）及 1.7.8 的 <worldbody> 合并修复 获得 viser 渲染能力（#1639）。viser.* 配置组（port、max_envs、 display_mode、env_idx）在 appo/sac/flashsac 配置上对齐 ppo 提供。

• Added the IsaacSim PhysX GPU contact buffer knobs isaacsim_gpu_max_rigid_contact_count and isaacsim_gpu_max_rigid_patch_count to EnvCfg, forwarded with fail-fast validation to the UniSim factory and omitted when unset so older unisim-core releases fail closed instead of seeing unknown kwargs (#1636). EnvCfg 新增 IsaacSim PhysX GPU 接触缓冲旋钮 isaacsim_gpu_max_rigid_contact_count 与 isaacsim_gpu_max_rigid_patch_count，以 fail-fast 校验转发给 UniSim factory；未设置时省略该键，旧版 unisim-core 不会收到未知 kwargs，而是 fail closed（#1636）。

• Added the training.log_interval key (default 1, no behavior change) and batched TensorBoard scalar logging for rsl_rl PPO: each iteration’s add_scalar calls become a single event record gated by the interval, removing the writer-queue stall on network filesystems (#1646, #1648). 新增 training.log_interval 配置键（默认 1，不改变行为）与 rsl_rl PPO 的 TensorBoard 标量批量写入：每次迭代的 add_scalar 调用合并为单条 event record 并按间隔落盘，消除网络文件系统上的写入队列阻塞（#1646、#1648）。

• Documented and tested the Motrix cpu_ids worker-affinity wiring, and added a durable A/B microbenchmark for the fused body-state read (fused packed read 5.7x, caller-owned-out copy 11.7x over the legacy cache-based copy at 8192 envs x 14 tracked bodies; end-to-end Motrix runs show no regression, so the single fused path is kept) (#962, #1308, #1634, #1635). 记录并测试 Motrix cpu_ids worker 亲和性接线，并为 fused body-state 读取新增持久化 A/B 微基准（8192 envs x 14 tracked bodies 下 fused packed read 快 5.7 倍、caller-owned-out copy 快 11.7 倍；端到端 Motrix 运行无回归，保留单一 fused 路径）（#962、#1308、#1634、#1635）。

Changed / 变更

• Raised the base UniSim requirement to unisim-core>=1.7.8 and mjbatch-uni to ~=0.2.4. Interactive rendering now consumes the unisim 1.7.5 physics-state playback contract through unilab.visualization.playback_state: snapshots split through the physics-state layout contract instead of hardcoded mjSTATE_FULLPHYSICS sizes, mocap geometry replays through get_playback_mocap_state, and playback entrypoints precheck supports_physics_state_playback and fail closed with a targeted message (#1637, #1639). 基础 UniSim 依赖提升到 unisim-core>=1.7.8，mjbatch-uni 提升到 ~=0.2.4。交互式渲染通过 unilab.visualization.playback_state 消费 unisim 1.7.5 的 physics-state 回放契约：快照按 physics-state 布局契约 拆分，不再硬编码 mjSTATE_FULLPHYSICS 尺寸；mocap 几何通过 get_playback_mocap_state 回放；回放入口预检 supports_physics_state_playback 并以明确信息 fail closed （#1637、#1639）。

• Bumped the optional uni_rl extra and dev environment from unilab-rl 1.3.2 to 1.4.3: 1.4.0 delivers WarpSAC, off-policy metric consumers migrated to the 1.4.1 canonical metric schema, the off-policy inference request timeout config was removed, and 1.4.3 stops the spurious “Learner stopped before inference tick” collector crash after normal off-policy training completion (#1640, #1643, #1645, #1655, #1658, #1659, #1662). 可选 uni_rl extra 与 dev 环境的 unilab-rl 从 1.3.2 升级到 1.4.3：1.4.0 提供 WarpSAC；off-policy 指标消费方迁移到 1.4.1 的规范指标 schema；移除 off-policy 推理请求超时配置；1.4.3 修复了 off-policy 训练正常结束后 collector 误报 “Learner stopped before inference tick” 崩溃的问题 （#1640、#1643、#1645、#1655、#1658、#1659、#1662）。

• Bumped torch to 2.14 and unified the CUDA wheels on cu130 on Linux and Windows; measured on Apple Silicon, FlashSAC end-to-end throughput improves by 35% (#1647). ONNX policy export now defaults to opset 18, matching the onnxscript function library torch 2.14’s dynamo exporter emits (#1654, #1656). torch 升级到 2.14，Linux 与 Windows 的 CUDA wheel 统一到 cu130；Apple Silicon 上实测 FlashSAC 端到端吞吐提升 35%（#1647）。ONNX 策略导出默认 opset 调整为 18，与 torch 2.14 dynamo 导出器生成的 onnxscript 函数库 匹配（#1654、#1656）。

• Enabled full-objective compile for FlashSAC (#1641), and pinned uv to 0.12.5 across the project, CI, and Docker (#1657). FlashSAC 启用 full-objective compile（#1641）；uv 在工程、CI 与 Docker 中统一固定到 0.12.5（#1657）。

Fixed / 修复

• The motion CSV loader preserves headerless frames (#1653). 动作 CSV 加载器保留无表头的帧（#1653）。

• macOS MuJoCo interactive eval now runs through mjpython (Cocoa main thread), and playback exits directly afterward to avoid the free-threaded CPython 3.13 interpreter-finalization deadlock (#1629, #1630). macOS 上 MuJoCo 交互式 eval 改经 mjpython 运行（Cocoa 主线程），回放 结束后直接退出，避免 free-threaded CPython 3.13 解释器收尾死锁 （#1629、#1630）。

• The viser viewer batches scene updates and shuts down cleanly on Ctrl-C (#1633). viser 查看器批量提交场景更新，并在 Ctrl-C 时干净退出（#1633）。

• Play artifact failures are decoupled: ONNX export and video rendering run under a shared nonfatal guard across the off-policy, APPO, and rsl-rl play paths, so one artifact failure no longer blocks the other (#1654, #1656). 回放产物失败解耦：ONNX 导出与视频渲染在 off-policy、APPO 与 rsl-rl 回放路径上统一走非致命保护，单个产物失败不再阻塞另一个 （#1654、#1656）。

• Restored the ROCm runtime dependencies and covered SAC cold-start inference compilation (#1650); the CI test job no longer rolls git-tracked robot scene XMLs back to the cached snapshot revision after restoring the asset cache (#1639). 恢复 ROCm 运行时依赖，并覆盖 SAC 冷启动推理编译（#1650）；CI 测试任务 恢复资产缓存后不再将 git 跟踪的机器人场景 XML 回滚到缓存快照版本 （#1639）。

1.3.2 (2026-09-22)

Breaking changes / 破坏性变更

• Removed the TD3 algorithm: the conf/td3 owner config tree, the train_td3.py entrypoint, the TD3 structured configs, and the TD3 dispatch branches in training, playback, and the CLI (#1625). The optional unilab-rl dependency moves to 1.3.2, the release that removed uni_rl.algos.fast_td3. Use SAC or FlashSAC for off-policy training. 移除 TD3 算法：conf/td3 owner 配置树、train_td3.py 入口、TD3 结构化 配置以及训练、回放与 CLI 中的 TD3 分发分支（#1625）。可选 unilab-rl 依赖升级到移除了 uni_rl.algos.fast_td3 的 1.3.2 版本。off-policy 训练 请改用 SAC 或 FlashSAC。

Added / 新增

• Added SAC owners for g1_motion_tracking on the Genesis, IsaacGym, Newton, and IsaacSim backends, with per-backend domain-randomization splits narrowed to each backend’s declared capability set (#1618, #1619). 为 g1_motion_tracking 新增 Genesis、IsaacGym、Newton、IsaacSim 后端的 SAC owner，并按后端拆分域随机化，收窄到各后端声明的 capability 集合 （#1618、#1619）。

• Added IsaacSim PhysX solver knobs to EnvCfg (isaacsim_solver_position_iteration_count, isaacsim_solver_velocity_iteration_count, isaacsim_bounce_threshold_velocity, isaacsim_contact_offset, isaacsim_rest_offset, isaacsim_max_depenetration_velocity), forwarded by env_backend_kwargs to the bounded, readback-validated solver configuration UniSim 1.7.x exposes on the IsaacSim backend; None keeps the PhysX scene defaults (#1617). EnvCfg 新增 IsaacSim PhysX 求解器旋钮 （isaacsim_solver_position_iteration_count、 isaacsim_solver_velocity_iteration_count、 isaacsim_bounce_threshold_velocity、isaacsim_contact_offset、 isaacsim_rest_offset、isaacsim_max_depenetration_velocity），由 env_backend_kwargs 转发至 UniSim 1.7.x 在 IsaacSim 后端暴露的有界、 带读回验证的求解器配置；None 保持 PhysX 场景默认值 (#1617）。

Changed / 变更

• PPO now drives upstream rsl-rl-lib directly through the UniLab-owned VecEnv adapter and distributed helpers; uni_rl (the unilab-rl package) is an optional extra needed only for APPO, off-policy algorithms, and multi-GPU data-parallel PPO launches (#1621, #1622). Single-process PPO training and playback work without uni_rl installed. PPO 改为通过 UniLab 自己的 VecEnv adapter 与分布式 helper 直接驱动上游 rsl-rl-lib；uni_rl（unilab-rl 包）变为可选 extra，仅 APPO、 off-policy 算法与多卡数据并行 PPO 启动需要（#1621、#1622）。单进程 PPO 训练与回放不再依赖 uni_rl。

• Raised the base UniSim requirement to unisim-core>=1.7.4 and pinned unilab-rl==1.3.2 for the optional uni_rl extra and the dev environment; the MuJoCo backend uses the exact tracked-body-state path from the executor. 基础 UniSim 依赖提升到 unisim-core>=1.7.4，可选 uni_rl extra 与 dev 环境固定 unilab-rl==1.3.2；MuJoCo 后端改用执行器的精确 tracked-body-state 路径。

1.3.1 (2026-09-20)

Added / 新增

• Add the M2 entity consumer: typed physical sources/variants, explicit logical-to-physical binding, selected entity reset transactions and per-variant state/control defaults. A registered primitive task validates the same pickleable EnvFactory with MuJoCo and IsaacSim. Standard and ROCm dependency profiles require the released unisim-core>=1.5.0 package, without a Git source override. Regression tests use the unified worker and construct the nonfirst-root fixture through the public factory. 新增 M2 实体消费层：物理源/variant 类型化、显式逻辑到物理绑定、局部实体 reset 事务和逐 variant 状态/控制默认值。注册 primitive task 验证同一可 pickle EnvFactory 在 MuJoCo/IsaacSim 的行为。标准与 ROCm 依赖配置要求已发布的 unisim-core>=1.5.0 包，不使用 Git source 覆盖。回归测试使用统一 worker，并通过公共 factory 构造非首 free root 场景。

• Remove quadratic selected-row lookup and unused full-batch snapshots from entity reset staging. Reject mapped logical roots that point at descendant bodies, keeping reads, defaults and writes aligned. A bounded A/B script records separate row-index and sparse-field effects without claiming simulation throughput. 删除实体 reset 暂存中的平方级行查找和无用全批状态快照。拒绝指向后代 body 的逻辑 root 绑定，使读取、默认值和写入保持一致。限定规模的 A/B 脚本分别记录行索引与稀疏字段的影响，不宣称仿真吞吐提升。

• Added IsaacGym fixed-variant protocol and real-runtime coverage. The deterministic worker mock validates and echoes the construction-time variant assignment, while the external Preview-4 slow lane realizes per-env actor asset selection and per-variant keyframes; public layout drift fails closed. IsaacGym fixed variants and the still-pending 600-variant support decision are documented on the backend page. 新增 IsaacGym fixed-variant 协议层与真实 runtime 覆盖。确定性 worker mock 校验并回显 construction-time variant assignment；外部 Preview 4 slow lane 验证逐环境 actor 资产选择与逐变体 keyframe；公共布局漂移 fail closed。 IsaacGym fixed variants 及仍待决策的 600 变体支持边界已写入后端文档。

• Added the IsaacGym fixed-variant scale benchmark. Each variant count runs in a fresh child process and records source generation, construction time, live worker RSS, and control/physics/env-step rates; results are written as a versioned JSON artifact for the #1579 support decision. 新增 IsaacGym fixed-variant 规模 benchmark。每个 variant 数在独立子进程中 运行，记录源生成、构造时间、worker 实时 RSS 与 control/physics/env-step 速率，并输出版本化 JSON artifact 供 #1579 support 决策使用。

Changed / 变更

• Moved the Motrix runtime pin into UniSim. The motrix extra now resolves unisim-core[motrix]>=1.7.3 instead of pinning motrixsim-core==0.8.2 directly, so the consumed Motrix runtime always matches the version the unisim.backend.motrix adapter is tested against (currently motrixsim-core 0.10.1). 将 Motrix runtime 的版本固定移入 UniSim。motrix extra 现在解析 unisim-core[motrix]>=1.7.3，不再直接固定 motrixsim-core==0.8.2， 使消费的 Motrix runtime 始终与 unisim.backend.motrix 适配层的测试 版本一致（当前为 motrixsim-core 0.10.1）。

• Raised the base and optional SuperDex UniSim requirements to unisim-core>=1.7.3. The released package adds IsaacSim mapped-scene per-body net contact force/found sensors, bounded PhysX solver configuration, per-entity self-collision and mapped reset domain randomization, plus IsaacGym mapped-scene reset randomization and interval body wrenches (UniSim #248/#249/#251/#255). 将基础与可选 SuperDex 的 UniSim 依赖提升到 unisim-core>=1.7.3。已发布 包新增 IsaacSim mapped scene 的逐 body 净接触力/found 传感器、有界 PhysX solver 配置、逐实体 self-collision 与 mapped reset 域随机化，以及 IsaacGym mapped scene 的 reset 域随机化和 interval body 力矩 （UniSim #248/#249/#251/#255）。

• Raised the base and optional SuperDex UniSim requirements to unisim-core>=1.7.2. The released package adds the SuperDex portable-scene physical-root integration, including physical kinematic roots, portable collision-disabled mirrors and scoped selected-control reset restoration (UniSim #154). 将基础与可选 SuperDex 的 UniSim 依赖提升到 unisim-core>=1.7.2。已发布 包新增 SuperDex portable scene 的物理 root 集成，包括物理 kinematic root、 portable collision-disabled mirror 和限定范围的 selected-control reset 恢复（UniSim #154）。

• Raised the base and optional SuperDex UniSim requirements to unisim-core>=1.7.1. The released packages contain portable MJCF profile v1 and the IsaacSim worker dependency-isolation fix, so consumer acceptance no longer uses the provisional local-checkout skip and validates the robot/object/table/mirror fixture from PyPI dependencies (#1609; UniSim #154/#155). 将基础与可选 SuperDex 的 UniSim 依赖提升到 unisim-core>=1.7.1。已发布 包包含 portable MJCF profile v1 与 IsaacSim worker 依赖隔离修复，因此 消费验收不再使用临时本地 checkout skip，并基于 PyPI 依赖验证 robot/object/table/mirror fixture（#1609；UniSim #154/#155）。

• Raised the base UniSim requirement to unisim-core>=1.4.1 to consume the published IsaacGym fixed-variant adapter. The optional superdex extra’s own >=1.4.0 constraint is unchanged: the base requirement already forces every installed profile to 1.4.1 or newer, and SuperDex has no 1.4.1-specific dependency change. The UniLab package version is unchanged. 将基础 UniSim 依赖提升到 unisim-core>=1.4.1，以消费已发布的 IsaacGym fixed-variant adapter。可选 superdex extra 自身的 >=1.4.0 约束保持不变： 基础依赖已经强制所有安装 profile 使用 1.4.1 或更新版本，且 SuperDex 在 1.4.1 中没有专属依赖变化。UniLab 包版本保持不变。

1.3.0 (2026-09-14)

Breaking changes / 破坏性变更

• The MuJoCo physics executor now uses the published mjbatch-uni~=0.2.0 package instead of mujoco-uni-runtime/mujoco_uni (#1552, #1553). The mujoco extra is aligned with MuJoCo 3.11 and checks for the mjbatch module at runtime. Chunk/forward-sizing knobs and the old Makefile bootstrap targets were removed because scheduling now belongs to mjbatch. Numerical equivalence with the previous executor is not guaranteed; the accepted Go2 drift is recorded by the #1554 baseline. The backend remains Linux/macOS only because mjbatch has no Windows wheels. MuJoCo 物理执行器改用已发布的 mjbatch-uni~=0.2.0，不再使用 mujoco-uni-runtime/mujoco_uni（#1552、#1553）。mujoco extra 对齐 MuJoCo 3.11，并在运行时检查 mjbatch 模块。chunk/forward 调整项和旧 Makefile 引导目标已删除，调度改由 mjbatch 负责。不保证与旧执行器数值 等价；可接受的 Go2 漂移由 #1554 基线记录。由于 mjbatch 没有 Windows wheel，该后端仍仅支持 Linux/macOS。

• The legacy UniLab DomainRandomization provider protocol is removed (#1563, #1567). Fixed model identity is now construction-time state; reset and interval randomization use backend-owned event terms, ResetStateTransaction, and the public UniSim plan contract. The unilab.dr namespace and provider-side compatibility surface are gone. 移除 UniLab 旧版 DomainRandomization provider 协议（#1563、#1567）。固定 模型身份改为 construction-time 状态；reset 与 interval 随机化使用后端拥有的 event term、ResetStateTransaction 和公开 UniSim plan contract。 unilab.dr namespace 与 provider 侧兼容面已删除。

Added / 新增

• Task-owned fixed model/tool variants can differ per environment and are materialized through the UniSim construction-time plan contract (#1568). The SimToolReal representative workload covers fixed-tool rollout parity, reset-time mass/inertia randomization, CPU/MJWarp parity, and one PPO learning iteration. task-owned fixed model/tool variant 现在可以按环境不同，并通过 UniSim construction-time plan contract 物化（#1568）。SimToolReal 代表性工作负载覆盖 fixed-tool rollout、reset-time 质量/惯量随机化、CPU/MJWarp parity 以及一次 PPO learning iteration。

Fixed / 修复

• Reward managers no longer emit duplicate Episode_Reward/* reset extras; the existing episode log remains the source for those metrics (#1570). Reward manager 不再在 reset extras 中重复输出 Episode_Reward/*；既有 episode log 仍是这些指标的数据源（#1570）。

Changed / 变更

• The MuJoCo executor dependency is pinned to mjbatch-uni~=0.2.1 and unisim-core moves to >=1.4.0 (unisim#71, mjbatch_uni#28). unisim-core 1.4.0 adds per-environment gravity reset, the cross-backend body wrench/torque contract, and per-substep callback wrenches through PreStepControlOutput; on the MuJoCo pre-step control path the executor’s split-substep sensor copyout now refreshes tracked body state at every substep boundary (about 14x faster than the previous host-side recompute, with bit-identical trajectories). Tests that pinned the previous executor callback protocol and Isaac worker diagnostic text were updated to the published contracts. MuJoCo 执行器依赖固定为 mjbatch-uni~=0.2.1，unisim-core 升级到 >=1.4.0（unisim#71、mjbatch_uni#28）。unisim-core 1.4.0 新增逐环境 重力 reset、跨后端 body wrench/torque 契约，以及通过 PreStepControlOutput 的逐子步 callback wrench；MuJoCo pre-step control 路径改用执行器的 split-substep 传感器增量拷出，在每个子步边界刷新 tracked body state（相比此前主机端重算约 14 倍加速，轨迹逐位一致）。 原先固定旧执行器 callback 协议与 Isaac worker 诊断文案的测试已更新到 已发布契约。

1.2.0 (2026-09-10)

Breaking changes / 破坏性变更

• Playback overlays now use typed per-environment DebugPrimitive values; the old extra_data_getter marker-array API is removed. 回放 overlay 改用类型化 per-environment DebugPrimitive；旧的 extra_data_getter marker-array API 已移除。

• Removed downstream-specific tasks. Go2 arm manipulation/HIM-PPO moved to legged-manipulation_unilab, and Sharpa in-hand/HORA moved to the dedicated sharpa_rl_unilab repository (#1547). Their task configs, assets, scripts, compatibility imports, and core-owned helpers were removed from UniLab. The historical 1.2.0 label is retained, but these removals are compatibility-breaking; consumers should use the coordinated external repository commits. 移除下游专属任务。 Go2 机械臂操作/HIM-PPO 迁移到 legged-manipulation_unilab， Sharpa in-hand/HORA 迁移到专用 sharpa_rl_unilab 仓库（#1547）。相关任务 配置、资产、脚本、兼容导入和核心内专属 helper 已从 UniLab 移除。历史版本号 1.2.0 保持不变，但这些移除具有兼容性破坏；使用方应固定配套的外部仓库提交。

Added / 新增

• SuperDex integration. Added the optional superdex extra, initial FR3 joint-target and Go2 joystick owners, native persistent CPU workers with rank-owned CPU partitions, and the FR3 bot asset download from Hugging Face (#1533, #1534). Published SuperDex wheels replace the temporary source build on CPython 3.12/3.13 Linux x86_64. FR3 playback is interactive-only because its .superdex_bot has no MJCF visual model; Go2 also supports recorded video playback. SuperDex 集成。 新增可选 superdex extra、首批 FR3 joint-target 与 Go2 joystick owner、带 rank-owned CPU partition 的原生持久 CPU worker，以及 来自 Hugging Face 的 FR3 bot 资产下载（#1533、#1534）。在 CPython 3.12/3.13 Linux x86_64 上，已发布的 SuperDex wheel 取代临时源码编译。FR3 资产 .superdex_bot 没有 MJCF visual model，因此只支持交互回放；Go2 另支持录制视频回放。

• Third-party owner contracts. Entity/reset transactions support additional geom and joint-parameter writes, explicitly bound fixed mocap bodies, and versioned training-progress export/import for downstream community packages such as Wuji. Playback uses typed DebugPrimitive overlays and exposes an embeddable SnapshotPlaybackSession for custom evaluation loops. 第三方 owner contract。 Entity/reset transaction 支持更多 geom 与 joint 参数写入、显式绑定的固定 mocap body，以及供 Wuji 等下游社区包使用的版本化 training-progress 导出/导入。回放改用类型化 DebugPrimitive overlay，并暴露可 嵌入自定义评估循环的 SnapshotPlaybackSession。

Changed / 变更

• Compatibility updates. Requires unisim-core>=1.2.0, pins unilab-rl==1.2.0, and updates the ROCm lock profile. This carries the shared backend, collector, and playback contracts used by the new owners. Because the SuperDex extra can also bring plain MuJoCo, sim=mujoco now checks the mujoco_uni runtime binding. 兼容性更新。 要求 unisim-core>=1.2.0，固定 unilab-rl==1.2.0，并更新 ROCm lock profile。这承载新 owner 所需的共享 backend、collector 与 playback contract。由于 SuperDex extra 也可能带入普通 MuJoCo，sim=mujoco 现在会检查 mujoco_uni runtime binding。

Fixed / 修复

• PPO action-standard-deviation logging now uses the current policy distribution, including state-dependent action scales. Nested play-profile merging also preserves unrelated fields and sibling observation terms when a partial override is applied. PPO action standard deviation 日志现在读取当前 policy distribution，可正确覆盖 state-dependent action scale。嵌套 play profile 合并时，局部 override 也会保留 未相关字段和 sibling observation term。

1.1.0 (2026-09-06)

• Added the Newton backend owner path and native ViewerGL playback, including the initial G1 walk PPO/SAC owners (#1509, #1511). 新增 Newton backend owner 路径与原生 ViewerGL 回放，包括首批 G1 walk PPO/SAC owner（#1509、#1511）。

• Fixed multi-GPU routing for spawned data-parallel collectors: rank-local backend devices are selected before CUDA/runtime initialization, including Genesis on non-zero physical GPUs and Newton’s rank-local device override. 修复 spawn data-parallel collector 的多 GPU 路由：在 CUDA/runtime 初始化前选择 rank-local backend 设备，包括非零物理 GPU 上的 Genesis 以及 Newton 的 rank-local device override。

• Added third-party task-package discovery through the unilab.tasks entry point group and generic interval domain-randomization dispatch (#1501, #1504). 新增通过 unilab.tasks entry-point group 发现第三方 task package 的能力， 以及通用 interval domain-randomization dispatch（#1501、#1504）。

• Updated the shared runtime contracts: unisim-core>=1.1.3, MuJoCo/MJWarp extras on the MuJoCo 3.11 line, mujoco-uni-runtime==0.5.0, and the matching ROCm profile. Go1 mesh assets also moved to the shared Hugging Face asset flow and out of the published wheel/sdist. 更新共享 runtime contract：unisim-core>=1.1.3、MuJoCo/MJWarp extra 对齐 MuJoCo 3.11、mujoco-uni-runtime==0.5.0 以及配套 ROCm profile。Go1 mesh 资产同时迁移到共享 Hugging Face asset 流程，并从发布的 wheel/sdist 移除。

• Expanded the bilingual backend evidence, installation, sim-to-sim, and project-rationale documentation. 扩展了 backend 证据、安装、sim-to-sim 与项目定位的双语文档。

1.0.0 (2026-09-04)

• Promoted the 0.1 architecture to the first stable package line: consume the production unilab-rl==1.0.0 release instead of the 0.2 TestPyPI line, broaden the exact Torch pins to ranges below 2.12 (with the Linux aarch64 floor at 2.9), and publish canonical project metadata and links. 将 0.1 架构提升为首个稳定包版本线：消费生产版 unilab-rl==1.0.0 而不是 0.2 TestPyPI 版本，将精确 Torch 钉版放宽到低于 2.12 的范围（Linux aarch64 下限为 2.9），并发布规范项目元数据与链接。

• Narrowed the packaged task surface: non-production MicroDuck and T800 task families moved to downstream ecosystem packages, and unused large motion files were removed. The manager-based runtime and production task owners introduced in 0.1.0 remain the supported API. 收窄打包任务面：非生产 MicroDuck 与 T800 task family 迁移到下游生态包， 并移除未使用的大 motion 文件。0.1.0 引入的 manager-based runtime 与生产 task owner 仍是受支持 API。

• Fixed event-curriculum velocity updates to re-read the live velocity range and fail closed when newly activated angular axes were not bound at construction. Reset terms can also read root poses already staged in the same transaction. 修复 event-curriculum velocity 更新：每次事件应用都会重新读取当前 velocity range；构造期未绑定的角度轴被激活时会 fail closed。reset term 也可以读取同一 transaction 中已 stage 的 root pose。

• Added tag-based trusted PyPI publishing with tag/version and successful-CI checks, distribution metadata validation, wheel installation, and an import smoke test. 新增基于 tag 的 PyPI trusted publishing，包含 tag/版本与成功 CI 校验、发行 元数据检查、wheel 安装和导入 smoke test。

0.1.0 (2026-09-04)

• First functional PyPI package. UniLab moved to the NumPy Manager-Based API with Hydra task owners for observations, actions, rewards, terminations, events, commands, curricula, and metrics; legacy monolithic environments were removed (#1042). 首个可实际使用的 PyPI 包。UniLab 迁移到 NumPy Manager-Based API，使用 Hydra task owner 声明 observation、action、reward、termination、event、command、 curriculum 与 metrics；旧的一体化 environment 已移除（#1042）。

• Physics adapters moved behind unisim-core>=0.1.14 (unisim) with MuJoCo, Motrix, MJWarp, Drake, Genesis, IsaacGym, and IsaacSim identities. RL algorithms, runners, collectors, IPC, and training logs moved behind unilab-rl==0.2.0 (uni_rl), preserving the package boundary that prevents learner code from importing UniLab. 物理适配器迁移到 unisim-core>=0.1.14（unisim），覆盖 MuJoCo、Motrix、 MJWarp、Drake、Genesis、IsaacGym 与 IsaacSim。RL 算法、runner、collector、 IPC 与训练日志迁移到 unilab-rl==0.2.0（uni_rl），保持 learner 代码不得 导入 UniLab 的包边界。

• The unified train, eval, and demo CLI covers PPO, APPO, SAC, TD3, and FlashSAC workflows; robot/scene/motion assets and demo checkpoints download from Hugging Face on first use. HORA and HIM-PPO remained script-level workflows at this release. 统一 train、eval 与 demo CLI 覆盖 PPO、APPO、SAC、TD3 与 FlashSAC 工作流；robot/scene/motion 资产和演示 checkpoint 首次使用时从 Hugging Face 下载。HORA 与 HIM-PPO 在该版本仍是脚本级工作流。

0.0.0 (2026-05-30)

• Placeholder release used only to reserve the unilab PyPI name. It contained no UniLab runtime code and declared no dependencies. 仅用于保留 unilab PyPI 名称的占位发布，不包含 UniLab runtime 代码，也没有 声明依赖。