Summary
ABot-M0.5 is a World Action Model built specifically for mobile manipulation—combining navigation and arm control within a single learned model. It identifies and addresses three fundamental bottlenecks in prior WAMs: (1) Temporal Granularity Mismatch (coarse video chunks obscure fine contact dynamics), (2) Action Structure Mismatch (entangled base and arm control spaces cause interference), and (3) Rollout Condition Mismatch (ground-truth training vs. autoregressive inference divergence causing error accumulation).
Key Contributions
- Identifies three structural mismatches in WAMs applied to mobile manipulation
- Fine-grained temporal representation for capturing contact-level dynamics
- Disentangled action space modeling for navigation vs. arm control
- Train-test consistency mechanisms to reduce error accumulation during autoregressive rollout
- Extends ABot series to mobile manipulation scenarios
Significance
First WAM designed explicitly for the combined challenges of mobile manipulation, offering a principled analysis of why general WAMs underperform in this setting and targeted solutions for each identified gap.