Skip to content

Event

Real-world evaluation of 281 imitation-learning policies for dexterous manipulation

1 reports · 1 sources

The research team used a five-fingered robotic hand to grasp and insert RJ45 connectors, evaluating 281 imitation-learning policies spanning seven policy families across approximately 1,700 real-world trials. Controlled ablations showed that observation and action representations had a greater impact than conventional optimizer and sampling choices: among point-cloud policies, region-of-interest cropping, robot-base-frame representation, and normalization within the point encoder each improved success rates by 70 to 100 percentage points; for vision-language-action policies, switching to absolute action prediction improved insertion success rates by up to 70 percentage points. Of 67 offline metrics, only 5 metrics predicting motion smoothness and directness were reliably correlated with insertion success (r = -0.70 to -0.80), and only within a single policy family. Skill decomposition, multi-view perception, and recovery demonstrations were associated with the best-performing systems, but the study did not isolate the effects of these factors individually.

Report timeline

  1. From Teleoperation to Long-Horizon Dexterity: Evaluating Visuomotor Skills and Offline Proxies for Dexterous Manipulation

    The research team used a five-fingered robotic hand to grasp and insert RJ45 connectors, evaluating 281 imitation-learning policies spanning seven policy families in approximately 1,700 real-robot trials. Controlled ablations showed that observation and action representations had a greater impact than conventional optimizer and sampling choices. For point-cloud policies, region-of-interest cropping, representation in the robot-base frame, and normalization within the point encoder each improved success by 70 to 100 percentage points. For the vision-language-action policy, switching to absolute action prediction improved insertion success by up to 70 percentage points.

    arXiv cs.RO ·

Related topics

Grasping and manipulationRobot models and learningResearch papersBenchmarks and evaluations

Interest trend

Current
3
Peak
4Nov 4 01:00
Past 24 hours
–

The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.