Reinforcement learning (RL) is increasingly used to post-train vision-language-action (VLA) models, but every update consumes robot rollouts that are slow and costly to collect, making sample efficiency a central concern. Manipulation tasks typically provide only sparse rewards, so a weak policy fails almost every rollout early in training and has little to learn from, even when those failures execute coherent behavior. Such a failure, however, is a success at a different task. We present Learning from Hindsight (LfH), which brings hindsight relabeling to RL post-training of VLAs by scoring failed rollouts against the tasks they actually achieved. A single vision-language model relabels both the instruction and the reward, proposing a hindsight instruction for a group of failed rollouts and scoring how well each satisfies it, and the policy trains on the relabeled and original rollouts jointly. Because VLAs generalize across language, relabeling in language lets the policy learn more from the same trajectories. On out-of-distribution LIBERO-PRO tasks, where standard RL improves only slowly, LfH achieves $5\times$ improvement in sample efficiency, and outperforms a dense progress-reward baseline. The gains hold across VLA backbones and on a physical Franka robot.
Iris Xu, Sunshine Jiang, John Marangola et al.· 0 citations
Mobile manipulators on construction sites offer considerable potential for increasing productivity, as the transport of materials and the execution of precise assembly work can be increasingly automated. However, the coordination of several such robots is a complex planning task, as task assignment, navigation, and reachability planning must be solved simultaneously and under dynamic environmental conditions. This work presents a multi-agent reinforcement learning (RL) approach that enables multiple mobile manipulators to complete a set of tasks in a structured environment. Each agent makes decentralized decisions about task selection and navigation, with kinematic reachability ensured by an integrated inverse kinematic solver. The policy is trained using proximal policy optimization (PPO), supported by a reward function that encourages both navigation progress and efficient task distribution. Simulation results show that the trained model is able to efficiently distribute tasks among multiple robots while taking kinematic constraints into account. The proposed method is superior to a greedy baseline that selects the nearest available task. With four robots and 35 tasks, the multi-agent RL approach achieves a success rate of 100%, while the baseline reaches only 55%.
Charlotte Stein, Yuheng Zhi, Michael C. Yip et al.· 2026 IEEE/ASME International...· 0 citations
