Event
RoboCode: Learning Reusable Action Primitives
1 reports · 1 sources
Xianyue Peng, Devi Amarsaikhan, and Iman Soltani propose RoboCode, which uses a shared soft-gated primitive action interface to learn short-horizon action patterns that can be reused across tasks: action chunks are represented by soft combinations of multiple learned primitives rather than a single discrete encoding. The interface is first learned on action chunks aggregated across multiple tasks, then frozen, and task-specific state encoders are trained to map the current state into the shared latent space. On five ManiSkill manipulation tasks, RoboCode achieves the highest or tied-highest closed-loop success rate, averaging 75.6%, higher than ACT's 68.8% and Diffusion Policy's 41.6%; the learned primitive interface reconstructs action chunks with an average MSE of 0.0187.
Report timeline
- RoboCode: Learning Reusable Action Primitives for Robot Manipulation
RoboCode learns short-horizon motion patterns that can be reused across tasks through a shared soft-gated primitive action interface, representing action chunks as soft compositions of multiple learned primitives rather than a single discrete code. The interface is learned on action chunks pooled across multiple tasks and then frozen, after which task-specific state encoders are trained to map into the shared latent space. On five ManiSkill tasks, it achieves an average closed-loop success rate of 75.6%, above ACT's 68.8% and Diffusion Policy's 41.6%, with an average action chunk reconstruction MSE of 0.0187.
arXiv cs.RO ·