Multimodal Failure in Action-Chunking Behavioral Cloning
A recent study published on arXiv (2605.22493) examines the shortcomings of behavioral cloning when a single observation can lead to multiple appropriate actions, concentrating on action-chunking strategies. Latent-variable policies encounter issues with posterior-prior regularization: excessive regularization eliminates crucial action-conditioned data necessary for mode differentiation, while insufficient regularization results in unreliable sampling if the prior fails to encompass significant latent areas. Additionally, action-space generative policies are limited by the smoothness of the transport from base to action; a mapping with a low Lipschitz constant struggles to assign high probability to distinct modes, necessitating abrupt transitions in the base space or off-support connections in the action space. Experiments conducted with synthetic data highlight these failure modes.
Key facts
- arXiv paper 2605.22493 studies multimodal failure in action-chunking behavioral cloning
- Behavioral cloning becomes difficult when the same observation admits several valid actions
- Latent-variable policies: posterior-prior regularization affects mode distinction and sampling reliability
- Action-space generative policies: multimodality constrained by Lipschitz constant of base-to-action transport
- Covering many modes requires sharp transitions in base space or off-support bridge regions
- Experiments were conducted on synthetic data
Entities
Institutions
- arXiv