Large discrete policy improves behavior modeling for autonomous systems
Large Discrete Policy: Advancing Explicit Behavior Modeling with Stochastic Iterative Scoring
RoboticsComputer Vision and Pattern Recognition
Summary
Models that control behavior, like in self-driving cars or robots, often use continuous approaches that can be hard to understand and sometimes give strange results. The authors introduce Large Discrete Policy (LDiP), which picks actions from a big set of realistic choices and improves decision making by scoring and narrowing down options step-by-step with some randomness. This makes the decisions easier to interpret and more reliable. Their approach works better or as well as current best methods in tasks like driving and robotic manipulation.
What this means in practice
- •For autonomous vehicle engineers: Enhance the reliability and interpretability of decision-making modules in self-driving cars by using discrete candidate actions scored iteratively.
- •For robotics developers: Improve robot manipulation policies by selecting from explicit discrete action sets with progressive scoring to achieve robust physical interactions.
Authors
Zhenxin Li, Nadine Chang, Xinglong Sun, Jingde Chen, Wenhao Yao, Zi Wang, Maying Shen, Yu-Gang Jiang, Zuxuan Wu, Shiyi Lan, Jose M. Alvarez
Abstract
Behavior policies are often formulated as continuous generative models, whose iterative denoising processes are expressive but difficult to interpret and prone to producing implausible actions. We propose the Large Discrete Policy (LDiP), a fully discrete behavior modeling framework that selects actions from a large vocabulary of physically plausible candidates. Rather than perturbing actions, LDiP improves expressivity through stochastic iterative scoring: it progressively re-scores and prunes candidates with score-space stochasticity, enabling fine-grained ranking and exploration among plausible actions while preserving an explicit decision process. Across end-to-end planning, closed-loop driving, robotic manipulation, and vision-language-action settings, LDiP consistently outperforms strong discrete and continuous baselines in autonomous driving, and exceeds or matches continuous generative policies in robotic manipulation. These results show that discrete policies, when equipped with effective scoring mechanisms, offer an expressive, plausible, and interpretable alternative for behavior modeling. Project website: https://zhenxinli.net/LargeDiscretePolicy/.