Robot manipulation improved by learning reusable action building blocks
RoboCode: Learning Reusable Action Primitives for Robot Manipulation
RoboticsMachine Learning
Summary
Robots often struggle by trying to plan their movements all at once without reusing parts that appear often. The authors developed RoboCode, which helps robots learn and reuse small action patterns seen in many tasks, like building blocks for motions. This approach makes robots better at performing a range of tasks, improving success rates compared to other methods. The learned building blocks are shared across tasks, making the robot's decisions more flexible and easier to understand.
What this means in practice
- •For robotics engineers: Enable robots to perform varied manipulation tasks by reusing learned short motion patterns to increase success rates and flexibility.
- •For automation system integrators: Incorporate a reusable action interface to build more efficient and interpretable robot control software for manufacturing and logistics.
Authors
Xianyue Peng, Devi Amarsaikhan, Iman Soltani
Abstract
Recent robot manipulation policies commonly predict continuous action chunks directly from observations, without explicitly modeling reusable structure in the action space. However, manipulation demonstrations often contain recurring short-horizon motion patterns that are shared across tasks. We introduce \textsc{RoboCode}, which learns these patterns through a shared soft-gated primitive action interface. Rather than assigning each action chunk to a single discrete code, \textsc{RoboCode} represents it as a soft composition of multiple learned primitives, providing a flexible intermediate representation of short-horizon actions. The primitive interface is learned from action chunks pooled across multiple tasks and then frozen while task-specific state encoders are trained to map current state inputs into the shared latent space. Experiments on five ManiSkill manipulation tasks address three questions. First, the learned primitive interface reconstructs action chunks with an average reconstruction MSE of 0.0187. Second, analysis of code activations reveals structured code reuse across tasks. Third, \textsc{RoboCode} achieves the highest or tied closed-loop success rate on all evaluated tasks, with an average success rate of 75.6%, compared with 68.8% for ACT and 41.6% for the larger Diffusion Policy baseline. These results suggest that soft-gated primitive codebooks provide an effective and interpretable intermediate action representation for robot manipulation.