CMU CoRL paper: grasp pretraining boosts dexterous tool-use success by 33.3 points over DP3
davheld · x · 2026-11-04
A Carnegie Mellon team presents a CoRL 2026 paper, From Grasps to Dexterity: Large-Scale Grasp Pretraining for Dexterous Manipulation:
- Question: can large dexterous grasp datasets support articulated tool use — acquiring a tool, maintaining contact, operating moving parts?
- Method: a 355k-trajectory grasp-pretraining dataset pretrains a low-level goal-conditioned controller, then fine-tuned on task demos, within a hierarchical imitation learning framework (high-level hand sub-goal prediction + low-level control)
- Introduces DexCraft, a simulation benchmark with 6 articulated tool-use tasks (spray bottle, lighter, stapler, etc.); 5 teleoperated demos augmented to 100 via MimicGen
- Results: beats end-to-end diffusion policies and from-scratch hierarchical policies in sim and real; +33.3 percentage points full-task success over DP3 in the real world
Takeaway: grasp datasets can serve as scalable pretraining data for contact-rich dexterous manipulation.
More from Embodied
- NYU team open-sources MABEL, an $8K mobile bimanual robot built in 2 weeks — RemiCadene · 2026-11-05
- Turning an AirPods case into a working mouse — TheMoonMidas · 2026-11-05
- Tsinghua, NYU and Northwestern unveil RoboDoctor, a training-free framework for fixing robot actions — jiqizhixin · 2026-11-05
- Robotics exec says deployments start in Singapore, aiming for everywhere from cafes to factories — micoolcho · 2026-11-05
- Gus Robotics watches its Microduck robot production line run for the first time in Shenzhen — mishig25 · 2026-11-05
- Smartglasses CV tells flower pickers what to harvest, cutting seasonal training to zero — nptacek · 2026-11-05