The statistical objective of cloning
Behavior cloning fits from demonstration pairs , often through negative log likelihood. Deterministic regression learns a location statistic such as the conditional mean. Averaging multiple valid modes can produce an action supported by none of them: left and right paths around an obstacle can average into a collision.
The issue is conditional multimodality. Representation, loss, and inference rules must agree.
Choosing an action representation
| Representation | Capability | Limitation |
|---|---|---|
| Continuous regression | One predicted action | Squared loss may average modes |
| Gaussian/mixture | Explicit probabilistic components | Restricted family and mixture complexity |
| Action bins | Probabilities over modes or intervals | Quantization and combinatorial size |
| Discrete mode plus residual | Coarse mode with local precision | Mode assignment and residual fitting |
| Diffusion/flow | Complex joint continuous distributions | Sampling cost and timing |
A bin can be a uniform numeric interval or a cluster center; clustered modes need not be ordered intervals. BeT-style classification-plus-residual modeling separates mode selection from local action refinement.
Closed-loop distribution shift
Cloning trains on expert visitation. Deployment errors can move the system into poorly covered states. Classical compounding-error bounds require a specified error event, horizon, and worst-case assumptions; is not a numerical prediction for every robot. Interactive corrections, DAgger, perturbation recovery, and broader state coverage can mitigate the shift.
Chunks and evaluation
Prediction horizon and executed horizon differ. Chunk consistency, refreshed observations, and inference latency jointly determine behavior. Evaluate success, recovery, continuity, and complete latency in addition to offline prediction loss.
Related reading: the diffusion supplement, ACT, and action interfaces.