Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly? - 2026-06-24 - Research Publication - Precise assembly is contact-rich and sparse-reward, making demonstrations costly and reinforcement learning from scratch difficult. Play2Perfect first trains a dexterous hand on diverse free-space object play, then… - source: arxiv.org
Universal Manipulation Exoskeleton: Learning Compliant Whole-body Policies with Real-time Torque Feedback - 2026-06-12 - Research Publication - Introduces a low-cost upper-limb exoskeleton that provides real-time haptic torque feedback while recording whole-arm configuration and joint torque signals. Retargets to multiple robot arms including OpenArm, Franka, and… - source: arxiv.org
Improving Robotic Generalist Policies via Flow Reversal Steering - 2026-06-11 - Research Publication - Uses flow reversal to refine coarse reference actions from humans or VLMs into precise in-distribution robot actions. Supports zero-shot control, fast distillation into smaller policies, and bootstrapped RL improvement.… - source: arxiv.org
Robots Need More than VLA and World Models - 2026-06-04 - Research Publication - Argues that robot policy scaling is bottlenecked by converting unstructured data into grounded supervision. Frames four missing interfaces: behavior autolabelling, embodiment retargeting, physics-grounded 3D world… - source: arxiv.org
EXPO-FT: Sample-Efficient Reinforcement Learning Finetuning for Vision-Language-Action Models - 2026-05-25 - Research Publication - Introduces RL finetuning for pretrained VLA policies using online interaction, Q-guided sampling, residual edit policies, and human corrections. Reports 30/30 success across evaluated real-robot manipulation tasks with… - source: pd-perry.github.io
SimToolReal: An Object-Centric Policy for Zero-Shot Dexterous Tool Manipulation - 2026-02-18 - Research Publication - The evaluation covered 120 physical rollouts across 24 tasks, 12 objects and six tool categories. The object-centric policy exceeded retargeting and fixed-grasp baselines by 37% while requiring no object- or task-specific… - source: arxiv.org