ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning
ADEPT is a large-scale reinforcement learning framework for learning sim-to-real transferable dexterity across high degree-of-freedom robot embodiments. It pretrains a dexterous policy on a generic object reposing task, then post-trains downstream policies with this pretrained behavior as a prior. The pretrained policy zero-shots the reposing phase of downstream tasks, but naïve RL fine-tuning rapidly degrades this capability. A stable post-training recipe combines behavior-cloning distillation, critic warm-up, and conservative on-policy updates. A joint-space Geometric Fabric mediates between the RL policy and the robot. Post-trained teachers are distilled into perceptive students that zero-shot sim-to-real transfer on two embodiments, including a 23 DoF robot.
ADEPT introduces a reinforcement learning framework for dexterous robot manipulation that pretrains on a generic object reposing task and then post-trains for downstream tasks. The approach enables learning new behaviors that are difficult to discover from scratch on multi-fingered robots and avoids relearning skills for each new task. The pretrained policy can zero-shot the reposing phase of downstream tasks, but naïve fine-tuning degrades this capability. The authors address this with a stable post-training recipe involving behavior-cloning distillation, critic warm-up, and conservative on-policy updates. A joint-space Geometric Fabric is used to safely exploit full kinematic dexterity. Post-trained teachers are distilled into perceptive students that achieve zero-shot sim-to-real transfer on two embodiments, including a 23 DoF robot.
The key technical contribution is a stable post-training recipe that preserves pretrained reposing skills while adapting to new tasks. The use of behavior-cloning distillation, critic warm-up, and conservative on-policy updates prevents catastrophic forgetting of the pretrained policy. The joint-space Geometric Fabric acts as a safety layer between the RL policy and the robot, enabling full kinematic dexterity without unsafe actions. Distillation into perceptive students enables zero-shot sim-to-real transfer, suggesting the learned representations are robust to embodiment and perception differences.
This work advances the feasibility of general-purpose dexterous manipulation by reducing the need for task-specific training from scratch. The ability to pretrain once and post-train for many tasks could lower the cost and time to deploy robotic manipulation in industrial and service settings. The sim-to-real transfer on multiple embodiments indicates potential for cross-platform skill reuse, which is valuable for robotics companies developing diverse hardware.
ADEPT could reduce the cost and time required to develop dexterous manipulation skills for robots by enabling skill transfer across tasks and embodiments. This may accelerate deployment of robotic automation in logistics, manufacturing, and service industries. The framework's ability to avoid relearning skills for each task could improve return on investment for robotics R&D.
Next signals to watch include whether ADEPT scales to more embodiments and tasks, whether the post-training recipe generalizes to other pretrained policies, and whether the sim-to-real transfer holds in unstructured real-world environments. Commercial adoption may follow if the framework reduces engineering effort for dexterous manipulation. Further research may explore combining ADEPT with foundation models for perception and language.