Imitation Learning Workflow

Collect expert demonstrations, train offline, evaluate, then deploy a compatible policy. SRB supports compatible tasks, datasets, and algorithms. Robomimic requires one environment and compatible observations and actions.

1. Collect demonstrations

Teleoperate the task and keep successful demonstrations. Follow the success rules.

srb agent collect --env peg_in_hole --num_episodes 50

The keyboard is the default. A 6-DoF SpaceMouse gives higher fidelity. An XR headset adds an immersive view, not input (Extended Reality).

srb agent collect --env peg_in_hole --teleop_device spacemouse

For collection defaults, success-only behavior, and dataset paths, see the srb agent collect reference.

2. Train an IL policy

Train offline; Isaac Sim is not needed. Pass --dataset. A fresh learn run directory does not find earlier collect/ data and errors if no dataset is supplied. Verified private HDF5 copies stay open until workers, handles, and logging close. Configured allocation limits are checked before provider reads. See limits, provenance, and --normalize_actions.

srb agent learn --algo robomimic_bc --env peg_in_hole \
  --dataset logs/peg_in_hole/collect/<timestamp>/datasets/demos_<timestamp>.hdf5

For the RNN variant, use robomimic_bc_rnn:

srb agent learn --algo robomimic_bc_rnn --env peg_in_hole \
  --dataset logs/peg_in_hole/collect/<timestamp>/datasets/demos_<timestamp>.hdf5

HDF5 stores per-step dones, terminated, and truncated. dones = terminated OR truncated marks episode boundaries. Value learners bootstrap through time-limit and collector-horizon truncations; they stop only on terminated. Robomimic bcq, cql, iql, td3_bc, and iris (BCQ value component) read dones and treat truncations as terminal.

3. Evaluate the policy

Evaluate a learn checkpoint:

srb agent eval --algo robomimic_bc --env peg_in_hole env.num_envs=1

Robomimic uses one environment:

• Collection raises robomimic collection requires num_envs == 1 if configured otherwise.
• Evaluation rejects num_envs != 1 with integration.robomimic.vector_env.unsupported.
• Offline learn creates no environment.

Evaluation checks task success at rollout end. If the task has no success signal, it warns and reports success_rate: null, never 0%.

For normalized actions, checkpoint norm_factor_min and norm_factor_max restore the environment range. The caller cannot override this range. If the checkpoint lacks a range but the run directory has logs/normalization_params.txt, evaluation errors and requires retraining. Otherwise, it loads unscaled actions.

Nested observations use slash-joined training keys.

4. Deploy to hardware

Use IL for sim-to-real with a compatible task, checkpoint, and hardware interface. Hardware operation is unvalidated. See Sim-to-Real Transfer for real_agent gen and real_agent eval.

5. Optional: iterate

Augment with policy rollouts

Collect more demonstrations with a partially trained policy (DAgger-style). Apply the same --success_only rules as in step 1.

srb agent collect --algo robomimic_bc --env peg_in_hole \
  --num_episodes 200

Rerun srb agent learn on merged HDF5 files. It accepts one --dataset for union training.

Combine with RL

IL may not generalize to unseen conditions. Incompatible checkpoint formats prevent direct RL warm-starting. See the reinforcement learning workflow:

• Train RL from scratch on the same task while IL collects demos.
• Resume interrupted RL with srb agent train --continue_training or a same-framework --model.
• Prefill skrl_sac, skrl_ddpg, or skrl_td3 replay from canonical export with --demo-dataset, --demo-max-transitions (metadata, limits, resume).
• Prefill Dreamer sequence replay from the canonical format. See continuous-action requirements, persistence, capacity, and source-bound resume.

Next: Sim-to-Real Transfer