protomotions.envs package

Tasks share BaseEnv and differ through configuration. Motion tracking, steering and path following use control components, not separate environment subclasses. Their experiment files wire controls to observation, reward, termination and action functions.

BaseEnv

protomotions.envs.base_env.env.BaseEnv receives an EnvConfig, a robot config, device, terrain, simulator, scene library and motion library. The training entry point constructs these dependencies; the environment initializes the simulator. Scene and motion libraries can be empty.

The runtime interface is:

obs, info = env.reset()
obs, rewards, dones, terminated, extras = env.step(actions)

obs is a dictionary of named tensors. The agent converts it to a TensorDict for the model. Actions have shape [num_envs, num_actions]. dones includes episode timeouts; terminated identifies terminal states for value bootstrapping. Call reset(env_ids) for environments whose episodes ended; the agent owns this rollout/reset sequence. get_obs() returns the current observations without stepping the simulation.

Component ownership

• control_components own mutable task state: reference poses, commands, goals and timers.

• observation_components map context tensors to named model inputs.

• reward_components map context tensors to reward terms.

• termination_components define training episode failure conditions.

• action_config maps policy outputs to simulator commands.

Observations, rewards and terminations use MdpComponent to bind pure tensor functions to EnvContext paths. Sensor estimates and clean simulator state are separate sources; each binding selects which it needs. Do not add noise inside a reward or observation function to duplicate sensor behavior.

Examples

Task wiring

Task	Stateful control	Experiment
Motion tracking	MimicControl	examples/experiments/mimic/mlp.py
Steering	SteeringControl	examples/experiments/steering/mlp.py
Path following	PathFollowerControl	examples/experiments/path_follower/mlp.py

See Configuration System for component bindings and config lifecycle, and Evaluation components for benchmark scoring and temporary environment snapshots. Evaluation failure criteria are configured separately from training terminations.