Bi-DexHands

A reinforcement-learning benchmark from a Peking University team, using two Shadow dexterous hands in Isaac Gym.

Bi-DexHands comes from Peking University's PKU-MARL team; its paper, “Towards Human-Level Bimanual Dexterous Manipulation with Reinforcement Learning,” appeared in the NeurIPS 2022 Datasets and Benchmarks track, and its code repository is named DexterousHands. It places two Shadow dexterous hands in Isaac Gym (NVIDIA's GPU-parallel simulator), with dozens of bimanual tasks and thousands of target objects, and the tasks were designed to correspond to different levels of human motor skill drawn from the cognitive-science literature. The paper reports over 30,000 frames per second on a single RTX 3090. The benchmark covers single-agent, multi-agent (treating each hand as a separate agent), offline, multi-task, and meta-reinforcement learning. It concludes that PPO-family algorithms can master simple tasks at roughly the level of a 48-month-old child, that multi-agent methods help on tasks requiring close two-hand coordination, but that existing algorithms mostly fail under multi-task and few-shot settings.

ExampleIn the paper, PPO-family algorithms learn tasks like catching a thrown object or opening a bottle; tasks requiring tight two-hand coordination, such as lifting a pot or stacking blocks, are learned more easily with multi-agent reinforcement learning.

Also called

DexterousHands

Related

Dexterous Manipulation · Bimanual Manipulation · Shadow Dexterous Hand · Isaac Gym · Multi-Agent Reinforcement Learning · Proximal Policy Optimization

Sources

Towards Human-Level Bimanual Dexterous Manipulation with Reinforcement Learning (arXiv 2206.08686)
PKU-MARL/DexterousHands (GitHub)

As of

2022-10