On the need to collect failure data for robots
The post discusses the importance of gathering failure data for robotic systems, referencing a recent episode on the topic. It highlights that gripper closure artifacts are expected because contact scenarios are underrepresented in learned simulations. The author suggests continuing to rely on MuJoCo for evaluation by replaying real teleoperation data, where joint tracking remains within approximately 1 degree, indicating known limitations. A potential future direction mentioned is combining learned models for rendering with physics engines for handling contact interactions.
The author relays a discussion on the necessity of collecting robot failure data, noting that current learned simulators lack contact data at gripper closure points and suggesting a hybrid approach combining learned rendering with physics engines for evaluation.
Source: Chris Paxton — X Robotics · Read original article ↗
Article text · Original source · English
On the need to collect failure data for robots - we had a recent @RoboPapers episode on this
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
joint tracking error
≈1
View original evidence
We still lean on MuJoCo for eval: replaying real SO-101 teleop, the joints track within ~1°, so at least we know where it’s lying. Maybe the two meet with learned models handling rendering and physics engines handling contact?
Open source S3MuJoCo usage for evaluation
≈1
View original evidence
We still lean on MuJoCo for eval: replaying real SO-101 teleop, the joints track within ~1°, so at least we know where it’s lying. Maybe the two meet with learned models handling rendering and physics engines handling contact?
Open source S3learned models handling rendering
≈1
View original evidence
We still lean on MuJoCo for eval: replaying real SO-101 teleop, the joints track within ~1°, so at least we know where it’s lying. Maybe the two meet with learned models handling rendering and physics engines handling contact?
Open source S3physics engines handling contact
≈1
View original evidence
We still lean on MuJoCo for eval: replaying real SO-101 teleop, the joints track within ~1°, so at least we know where it’s lying. Maybe the two meet with learned models handling rendering and physics engines handling contact?
Open source S3
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
We still lean on MuJoCo for eval: replaying real SO-101 teleop, the joints track within ~1°, so at least we know where it’s lying. Maybe the two meet with learned models handling rendering and physics engines handling contact?
Open source S3
On the need to collect failure data for robots - we had a recent @RoboPapers episode on this
Open source S1
【引用 @srigopal】The gripper closure artifacts make sense, since contact is exactly where learned sims have the least data.
Open source S2
The gripper closure artifacts make sense, since contact is exactly where learned sims have the least data. We still lean on MuJoCo for eval: replaying real SO-101 teleop, the joints track within ~1°, so at least we know where it’s lying. Maybe the two meet with learned models handling rendering and physics engines handling contact? https://x.com/srigopal/status/2116940192745709913?a=bhBeXK9B7RgBm97zSIk&s=20View the quoted post on X
Source：Chris Paxton · x.com