ExoStart:

Efficient learning for dexterous manipulation with sensorized exoskeleton demonstrations

Overview

We present ExoStart, a general and scalable learning framework that leverages the power of human dexterity for robotic hand control. In particular, we obtain high-quality data by collecting direct demonstrations without a robot in the loop using a sensorized low-cost wearable exoskeleton, capturing the rich and nuanced behaviors that humans naturally demonstrate with their own hands. We also propose a simulation-based dynamics filter that generates dynamically feasible trajectories from the collected demonstrations and use the generated trajectories to bootstrap an auto-curriculum reinforcement learning method that relies only on simple sparse rewards. The ExoStart pipeline is generalizable and yields robust policies that transfer zero-shot on real robots. Our results demonstrate that ExoStart can generate dexterous real-world hand skills, achieving a success rate above 50% on a wide range of complex tasks such as opening an AirPods case or inserting and turning a key in a lock.

Supplemental video

exostart-website-version.mov

System pipeline

We present ExoStart, as a reliable and robust recipe for learning policies for real-world autonomous dexterous manipulation. Our framework employs a real-to-sim-to-real pipeline, starting with the collection of direct human demonstrations using a sensorized exoskeleton. These demonstrations are then used to generate dynamically feasible trajectories through the use of a simulation-based dynamics filter. Subsequently, an auto-curriculum reinforcement learning algorithm, driven by simple binary rewards, is bootstrapped using these trajectories. Finally, the learned policies are distilled and transferred from simulation to the real world in a zero-shot manner.

Results

Table 1 presents 1) the number of demonstrations collected with the exoskeleton in the real world, 2) the number of filtered trajectories after applying dynamics filtering, 3) the performance of the feature-based teacher policies in simulation, 4) the performance of our final vision-based student policies in solving the real-world tasks, as well as 5) the time limit for each episode during real-world evaluation.