DiscreteRTC: Discrete Diffusion Policies are Natural Asynchronous Executors

Researchers from UC Berkeley and other institutions have introduced DiscreteRTC, a novel approach to robotic control that leverages discrete diffusion policies for asynchronous execution. Unlike traditional synchronous executors that pause between action chunks—fatal for dynamic tasks—DiscreteRTC enables robots to think while acting through real-time chunking (RTC). The key insight is that discrete diffusion, which generates actions by iteratively unmasking tokens, naturally supports inpainting as a core operation rather than requiring external corrections.

This structural advantage eliminates the need for fine-tuning or heuristic guidance, reducing inference latency and computational overhead. In tests on dynamic simulated benchmarks and real-world manipulation tasks, DiscreteRTC achieved a 50% higher success rate on a dynamic pick task compared to flow-matching-based RTC, while using only 0.7x the computation of generating actions from scratch. The method requires zero lines of additional code for asynchronous inpainting, making it simpler and more efficient than prior continuous approaches.