Event
TRP: Behavior Cloning Policy with Recursive Latent State Refinement
1 reports · 1 sources
Hakan Aktas and Cengiz Oztireli propose Tiny Recursive Policy (TRP), a parameter-efficient behavior cloning policy: at each control step, it repeatedly updates high-level and low-level latent states using a shared computation module, then decodes the action sequence and passes the state to subsequent predictions. The authors evaluate TRP on three benchmarks: LIBERO-100, MetaWorld ML45, and LIBERO-Mem. They state that perturbation experiments during execution show that reliance on the persistently passed state increases with the task's need for temporal context; adding perturbation to the passed state during training markedly improves robustness to corruption of the passed state and to visual corruption. These effects come from experimental results reported by the authors themselves.
Report timeline
- Tiny Recursive Policy: Recursive Latent Refinement for Visuomotor Control
The authors propose Tiny Recursive Policy (TRP), a parameter-efficient behavior cloning policy that repeatedly updates high-level and low-level latent states using a shared computation module at each control step before decoding an action sequence, and propagates the state to subsequent predictions.
arXiv cs.RO ·