From Teleoperation to Long-Horizon Dexterity: Evaluating Visuomotor Skills and Offline Proxies for Dexterous Manipulation
From Teleoperation to Long-Horizon Dexterity: Evaluating Visuomotor Skills and Offline Proxies for Dexterous Manipulation
This research evaluates 281 imitation-learning policies across seven families using approximately 1,700 real-robot trials of RJ45 connector grasping and insertion. The study finds that observation and action representations dominate performance, with specific improvements like region-of-interest cropping and absolute action prediction yielding significant gains. Only five offline metrics reliably correlated with success within a single policy family, indicating that while offline metrics can shortlist policies, real-robot evaluation remains necessary.
The study evaluates 281 imitation-learning policies across seven families using approximately 1,700 real-robot trials of RJ45 connector grasping and insertion, finding that observation and action representations dominate performance while only five offline metrics reliably correlate with success within a single policy family.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
trials
≈1,700
Reported trials
View original evidence
hoices that govern real-world performance, 281 imitation-learning policies spanning seven policy families were evaluated in approximately 1,700 real-robot trials of RJ45 connector grasping and insertion using a five-fingered robotic hand. Controlled ablations showed that observation and action representations dominated conventional optimizer and sampling choices. For point-cloud policies, region-of-interest cropping,
Open source S3offline_metrics_evaluated
67
View original evidence
r effects were not independently isolated. Of 67 offline metrics evaluated, only five measures of predicted-motion smoothness and directness reliably correlated with insertion success (r = -0.70 to -0.80), and only within a single policy family. These results identify representation choices as key determinants of real-world imitation-learning performance and show that offline metrics can shortlist policies, but canno
Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
hoices that govern real-world performance, 281 imitation-learning policies spanning seven policy families were evaluated in approximately 1,700 real-robot trials of RJ45 connector grasping and insertion using a five-fingered robotic hand. Controlled ablations showed that observation and action representations dominated conventional optimizer and sampling choices. For point-cloud policies, region-of-interest cropping,
Open source S3
representation in the robot-base frame, and normalization within the point encoder each improved success by 70-100 percentage points. For the vision-language-action policy, absolute rather than relative action prediction improved insertion success by up to 70 percentage points. Skill decomposition, multi-view perception, and recovery demonstrations were associated with the strongest-performing systems, although thei
Open source S4
r effects were not independently isolated. Of 67 offline metrics evaluated, only five measures of predicted-motion smoothness and directness reliably correlated with insertion success (r = -0.70 to -0.80), and only within a single policy family. These results identify representation choices as key determinants of real-world imitation-learning performance and show that offline metrics can shortlist policies, but canno
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org