VOMMI: Collecting and Leveraging Portable Demonstrations for Mobile Manipulation
VOMMI: Collecting and Leveraging Portable Demonstrations for Mobile Manipulation
This research presents VOMMI, a framework for collecting portable demonstrations to address data scarcity in embodied intelligence. It connects RGB observations to vision-language-action (VLA) post-training through offline trajectory reconstruction and online visual-motion conditioning. The system synchronizes body and hand views without kinematic calibration, using R2-VO to refine trajectories with sparse geometric anchors. Experiments show that policies post-trained on portable demonstrations achieve 18.2% lower base-velocity error than those trained with robot-collected demonstrations, while maintaining comparable end-effector accuracy. Offline reconstruction reduces absolute trajectory errors by 24.6% on average, and the complete system improves mean success rate by 8.3 percentage points over OpenPI 0.5 across three real-robot tasks.
The paper introduces VOMMI, a framework that enables the collection of portable RGB demonstrations for mobile manipulation tasks without requiring robot hardware or kinematic calibration, achieving lower base-velocity error and improved success rates compared to baselines trained on robot-collected data.
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
trajectories
500
View original evidence
y into the base branch. Experiments use a 500-trajectory portable for each task, with 75 trajectories held out for RGB-VO evaluation, and 200 robot demonstrations as references. Our policy, post-trained only on portable demonstrations, achieves 18.2% lower base-velocity error than a policy trained with robot-collected demonstrations, while maintaining comparable end-effector translation accuracy. Offline reconstructi
Open source S5Robot trajectories
75 trajectories
Trajectory count
View original evidence
y into the base branch. Experiments use a 500-trajectory portable for each task, with 75 trajectories held out for RGB-VO evaluation, and 200 robot demonstrations as references. Our policy, post-trained only on portable demonstrations, achieves 18.2% lower base-velocity error than a policy trained with robot-collected demonstrations, while maintaining comparable end-effector translation accuracy. Offline reconstructi
Open source S5Robot count
200 robots
View original evidence
y into the base branch. Experiments use a 500-trajectory portable for each task, with 75 trajectories held out for RGB-VO evaluation, and 200 robot demonstrations as references. Our policy, post-trained only on portable demonstrations, achieves 18.2% lower base-velocity error than a policy trained with robot-collected demonstrations, while maintaining comparable end-effector translation accuracy. Offline reconstructi
Open source S5percent
18.2
View original evidence
y into the base branch. Experiments use a 500-trajectory portable for each task, with 75 trajectories held out for RGB-VO evaluation, and 200 robot demonstrations as references. Our policy, post-trained only on portable demonstrations, achieves 18.2% lower base-velocity error than a policy trained with robot-collected demonstrations, while maintaining comparable end-effector translation accuracy. Offline reconstructi
Open source S5percent
24.6
View original evidence
on reduces absolute trajectory errors for the body and hand streams by 24.6% on average relative to the best evaluated baseline for each stream. The complete system improves the mean success rate by 8.3 percentage points over OpenPI 0.5 across three real-robot tasks.
Open source S6percent
8.3
View original evidence
on reduces absolute trajectory errors for the body and hand streams by 24.6% on average relative to the best evaluated baseline for each stream. The complete system improves the mean success rate by 8.3 percentage points over OpenPI 0.5 across three real-robot tasks.
Open source S6
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
y into the base branch. Experiments use a 500-trajectory portable for each task, with 75 trajectories held out for RGB-VO evaluation, and 200 robot demonstrations as references. Our policy, post-trained only on portable demonstrations, achieves 18.2% lower base-velocity error than a policy trained with robot-collected demonstrations, while maintaining comparable end-effector translation accuracy. Offline reconstructi
Open source S5
on reduces absolute trajectory errors for the body and hand streams by 24.6% on average relative to the best evaluated baseline for each stream. The complete system improves the mean success rate by 8.3 percentage points over OpenPI 0.5 across three real-robot tasks.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org