pengyue-polaron/nyush-galaxea-a1-lingbot-va-plug-insertion-evaluations

LingBot-VA on Galaxea A1: plug-insertion Evaluations

Real-robot closed-loop evaluation of a LingBot-VA checkpoint trained to pick up a charger and insert it into the leftmost socket of a power strip.

Result

Five additional runs stopped at the initial grasp and are excluded because they did not evaluate insertion.

Representative rollouts

<table> <tr><th>Success</th><th>Failure</th></tr> <tr> <td><img width="100%" src="https://huggingface.co/datasets/pengyue-polaron/nyush-galaxea-a1-lingbot-va-plug-insertion-evaluations/resolve/main/assets/pluginsertionsuccess2x.gif"></td> <td><img width="100%" src="https://huggingface.co/datasets/pengyue-polaron/nyush-galaxea-a1-lingbot-va-plug-insertion-evaluations/resolve/main/assets/pluginsertionfailure2x.gif"></td> </tr> <tr> <td>Stable insertion followed by gripper release.</td> <td>Pins remain offset from the holes; no insertion.</td> </tr> </table>

Both GIFs show the final approach from the wrist camera at 2× speed.

Failure analysis

The 11 failures fall into two visible groups:

• —6/11: the plug reaches the power strip but residual lateral or orientation error prevents the pins from entering.
• —5/11: the failed approach is followed by a drop, detachment, or incorrect release.

Final alignment

<table> <tr><th>Failed final approach</th><th>Successful result</th></tr> <tr> <td><img width="100%" src="https://huggingface.co/datasets/pengyue-polaron/nyush-galaxea-a1-lingbot-va-plug-insertion-evaluations/resolve/main/assets/failedalignmenttargetvisible.png"></td> <td><img width="100%" src="https://huggingface.co/datasets/pengyue-polaron/nyush-galaxea-a1-lingbot-va-plug-insertion-evaluations/resolve/main/assets/successafter_release.png"></td> </tr> <tr> <td>The upper target holes remain visible, but the pins do not align with them.</td> <td>The only successful rollout after release.</td> </tr> </table>

The intended target is the upper socket opening, not the lower opening. It is visible in the wrist stream throughout the final approach, so the videos do not support camera occlusion as a failure cause. The supported conclusion is limited to a final position or orientation alignment error; these recordings do not establish its deeper control cause.

Evaluation details

• —Date: 2026-08-08
• —Prompt: pick up the charger and insert it into the first socket from the left on the power strip
• —Checkpoint: step-500
• —Training data: 31 A1 EEF demonstrations

Data

• —Evaluation manifest
• —All rollout recordings and logs