4.0 KiB
Platform design
One model, explicit contracts
Onboarding inspects input model files, mesh references, joint types and limits, actuation, and optional policy metadata. It recommends a route using declared intent and model evidence. The intake inventories upstream training configs separately from unified policy metadata. Metadata extraction reads JSON and YAML as data and Python as static syntax. It records provenance, conflicts and unresolved expressions for review. A native verification report records the model and compiled dependencies, controller inputs, and measured physics and response separately.
Tier 1 converts the source actor to a validated ONNX graph and canonical layout using explicit feature and joint order. Numerical parity tests check the adapter against the source graph. Tier 2 binds a chosen deterministic controller to named model elements. Both converge on the native verification and packaging path.
Locked training recipes
Tier 3 separates robot data from learning choices. A literal
tasks/<robot_id>/robot_spec.py declares the robot family, model and controls
paths, base body and height, contacts, and actuator calibrations. The registry
reads its literals without executing the module, and imports the selected
template only when building an environment.
The four recipes define body-frame velocity tracking with named proprioceptive actor channels and a privileged critic. Channel widths follow the model's selected joints and supports. Domain randomization changes support friction, base center of mass, and encoder bias, and adds pushes. The curriculum widens the command range after sustained tracking evidence, with the earned level and evidence saved in each checkpoint. Wheeled bipeds use joint position targets for legs and velocity targets for continuous wheel joints. Wheel rotation is excluded from the position observation. Native policy layouts record separate joint selections and the target kind per action channel.
The lock test protects family files and shared behavior dependencies. A new calibration changes only the robot specification.
One environment on two backends
mjlab's manager construction asks for a registered simulation class. Native MuJoCo implements the same batched interface that MuJoCo Warp exposes, so action, command, observation, reward, and PPO code is shared. Select the backend before constructing the environment. Native copies a model per environment and uses a thread pool. Warp batches state on the GPU. Backend and environment count affect the experience collected in a run and are part of any training report.
Backend validation
Sharing the environment interface does not imply identical contact solutions or learning curves. The example reports compare fixed-command checkpoint replay on Warp and native MuJoCo, and compare native checkpoint replay with exported ONNX. Each report records its own model, duration, command and noise setting. These checks establish measured behavior for those runs, not bitwise equality between simulators or equal training outcomes across different batch sizes. Whether long-rollout state differences remain bounded, how they vary with contact depth, and their effect on learning are not yet settled for this robot collection. Settling these questions requires paired state measurements and comparing learning curves with matched models, hyperparameters and seeds.
Warp records constraint/contact allocation alongside checkpoints and restores it for replay and export. An overflow stops collection before the affected observations reach PPO. A completed update with an overflow warning is not valid training evidence.
Packaging
The native app stores the model and referenced assets, exact mode artifacts, verification evidence, manuals, and a Python controller entry point with its host requirements. Model-only viewer apps verify compilation. Controlled modes require passing reports for the exact model, controller, and bindings being packaged. Legacy Rust artifacts are assembled by the existing bundle route. See the bundle format for consumer details.