Recent advances in frontier models have pushed the boundaries of agentic robot control, enabling agents to perform complex robotic tasks through reasoning and interaction. However, repeated calls to these models during execution can substantially increase task completion time. Code as Policies (CaP) offers a route for coding agents to acquire physical skills through experience and retain them in executable programs, allowing the reasoning invested in development to be reused across runs. Developing reliable code policies while limiting the cost of physical experimentation remains a central challenge. We investigate human guidance as a source of priors for feasible actions and simulation as an auxiliary environment for exploring and testing candidate policies. Our findings favor robot demonstrations available on request and simulation used selectively during real-world policy refinement. These findings inform ACE-CaP-0, an agentic recipe that couples policy revision with experience acquisition through Demonstrations on Request and Simulation on Demand. Across eight real-world manipulation tasks spanning alignment, multi-object handling, and bimanual coordination, the resulting frozen programs achieve 92.1\% success (221/240 trials). Mean execution times across tasks range from 0.97 to 2.27 minutes. The programs execute with runtime perception and control, without a human or coding agent in the loop. These results highlight CaP's potential to turn experience into reliable, reusable physical skills.
Linwei Dong, Ruoyu Guo, Ge Bai, Quan Zheng, Yawei Luo, Changqing Zou
原文摘要
Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-step models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.