Hardware Safety Layers

The policy produces actions under the training contract. A deploy-side safety layer must live between the policy output and the motor driver and reject contract violations before they become actuator commands.

Required components

Layer

Responsibility

Schema check

Action has correct dtype, shape, finite values. Reject NaN / Inf.

Range clamp

Clamp each joint target to deploy-configured joint limits.

Δ clamp

Reject or clamp per-step action deltas using a threshold owned by the deploy controller.

Rate limit

Slew-rate limit applied AFTER clamp.

Watchdog

If no fresh action arrives within the controller-owned timeout, hold the last known safe target or enter the controller’s safe state.

Pose monitor

Roll / pitch outside operating envelope → triggered fault.

Operator stop

Big red button → instant torque-disable, regardless of state.

Where the safety layer lives

        flowchart LR
    P[Policy ONNX] --> S[Safety layer<br/>C++ on robot computer]
    S -->|safe target| D[Motor driver]
    D -->|encoder + IMU| Pre[Observation builder]
    Pre --> P
    S -.->|fault| OP[Operator UI]
    OP -.->|E-stop| D

Keep the hard real-time safety checks in the deploy controller, not in the training script. The repository does not implement a production motor-driver safety loop — that boundary is yours to build and test.

What the policy assumes you’ve configured

The policy expects the action mapping and limits its training owner declared. For the G1 WBT owner (src/unilab/conf/sac/task/g1_wbt_obs/mujoco.yaml):

Quantity

Authority

action scale

env.actions.joint_pos.scale (scalar 2.0 for this owner; other owners declare a regex → value map resolved per actuator)

default angles

use_default_offset: true, i.e. the stand keyframe joint block of the owner’s scene XML

joint limits

jnt_range in the scene XML

kp / kd

position actuator gainprm / biasprm in the scene XML

Derive these from the owner YAML and its scene, and reproduce the owner’s resolved per-actuator scale exactly — a scalar owner and a regex-map owner are not interchangeable. Do not hand-copy joint ranges or gains into a second place that can silently drift from the asset.

Hand-off testing

Before integrating policy → safety → motor, test the safety layer in isolation:

  1. Inject a NaN action and verify the command is rejected.

  2. Inject an out-of-range joint target and verify clamping uses the joint range from the owner’s scene XML.

  3. Cut the policy feed mid-run and verify the controller enters its configured safe state.

See also