Imitation Learning Workflow

Imitation learning (IL) collects demonstrations with collect and trains policies offline with learn. Support depends on the task, dataset, and algorithm. Robomimic requires one environment and compatible observation and action data.

1. Collect Demonstrations

Reference: srb agent collect — Collect Demonstrations

Teleoperate the task and discard failed attempts; keyboard save does not mark a demo successful. --success_only keeps demos only when the ending episode’s success flag or configured success termination term reports success (see collection tips):

srb agent collect --env peg_in_hole --num_episodes 50

Keyboard is the default input. For 6-DoF input, use a SpaceMouse. An XR headset provides an immersive viewpoint; see XR headset guidance:

srb agent collect --env peg_in_hole --teleop_device spacemouse

Datasets default to logs/<env_id>/collect/<timestamp>/datasets/demos_<timestamp>.hdf5. Set --logdir to change the root.

2. Train an IL Policy

Reference: srb agent learn — Learn from Demonstrations

See the srb agent learn reference for its execution requirements.

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

Pass --dataset explicitly; learn creates a fresh run and does not discover data from earlier collect runs. The RNN variant uses the same workflow:

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

Training verifies private dataset copies and checks allocation limits before reading them. See dataset admission and normalization.

HDF5 demos record dones, terminated, and truncated. Robomimic value learners use dones for episode boundaries and currently treat truncations as terminal.

3. Evaluate the Policy

Reference: srb agent eval — Evaluate Agent

Evaluate with srb agent eval:

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

Robomimic uses one environment. Evaluation and collection reject num_envs != 1; offline learn does not create an environment.

• Success uses the task’s episode success signal; tasks without one report success_rate: null and a warning.
• When training uses --normalize_actions, the checkpoint stores the action range used at evaluation.
• Older checkpoints with logs/normalization_params.txt require retraining; otherwise their normalization cannot be detected.
• If no action range is detectable in checkpoint metadata, evaluation passes normalized policy actions to the environment without rescaling. Nested observations use the same slash-joined keys as training.

4. (Optional) Iterate

Augment with Policy Rollouts

Use a partially trained policy to collect demonstrations:

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

Merge the HDF5 files before training on multiple collection runs; learn accepts one --dataset.

Combine with RL

IL checkpoints cannot warm-start RL because frameworks use different formats.

• Collect demonstrations with IL while training RL from scratch.
• Resume an RL run with --continue_training or a same-framework --model.
• Prefill SKRL or Dreamer replay from canonical demonstration exports; see SKRL and Dreamer requirements.

5. Deploy to Hardware

Reference: Sim-to-Real Transfer

Use the sim-to-real workflow when task, checkpoint, and hardware interfaces are compatible; hardware operation remains gated.

See Also

• Imitation Learning Integration — framework details
• Reinforcement Learning Workflow — complementary paradigm
• Extended Reality — high-fidelity demonstration source