Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied Manipulation

While current embodied policies exhibit remarkable manipulation skills, their execution remains unsatisfactorily slow as they inherit the tardy pacing of human demonstrations. Existing acceleration methods typically require policy retraining or costly online interactions, limiting their scalability for large-scale foundation models. In this paper, we propose **S**peed**u**p **P**atch (**SuP**), a lightweight, policy-agnostic framework that enables **plug-and-play acceleration** using solely offline data. SuP introduces an external scheduler that adaptively downsamples action chunks provided by embodied policies to eliminate redundancies. Specifically, we formalize the optimization of our scheduler as a Constrained Markov Decision Process (CMDP) aimed at maximizing efficiency without compromising task performance. Since direct success evaluation is infeasible in offline settings, SuP introduces **World Model based state deviation** as a surrogate metric to enforce safety constraints. By leveraging a learned world model as a virtual evaluator to predict counterfactual trajectories, the scheduler can be optimized via offline reinforcement learning. Empirical results on simulation benchmarks (Libero, Bigym) and real-world tasks validate that SuP achieves an overall $1.8\times$ execution speedup for diverse policies while maintaining their original success rates.

Robot-control AI systems can now follow instructions and perform complex manipulation tasks, but many still move slowly because they copy the cautious pace of human demonstrations. A simple shortcut, such as skipping every other action, can make a robot faster, but it can also make the robot drift away from the motion needed to succeed. We introduce Speedup Patch (SuP), a lightweight add-on that speeds up an existing robot policy without retraining it or testing risky new behaviors on real hardware. SuP looks at the short action sequence the original policy plans to execute and chooses when it is safe to skip redundant intermediate motions. It learns this decision from recorded demonstrations by training a small predictor of the robot's motion, then using that predictor to check whether faster action sequences stay close to the original path. In simulated benchmarks and real robot tasks, SuP made diverse policies run about 1.8 times faster overall while preserving their success rates. This shows that many robot policies contain unused speed that can be unlocked with a practical plug-in rather than rebuilding the policy from scratch.