Robotic hand self-righting with visual feedback

I gave GPT-6 Astra (Ultra) access to a Wuji2 hand and a third-person camera, and asked it to get the hand upright. Watching it use its fingers to lift itself off the table felt a little like watching a baby learn to stand. The hand stayed upright after the motors were disabled.

Ultra succeeded in about 3 minutes. An earlier run using Max Effort failed after about 8 minutes of repeated attempts and adjustments.

One prompt

We have connected a third-person camera and a Wuji2 robotic hand for control. Please control the robotic hand to move itself into an upright configuration, with its base positioned at the bottom, rather than lying flat as it is now. You need to figure out on your own how to adjust the robotic hand pose to achieve this.

Please record a video while operating.

How Astra produced the motion

Astra used Wuji's existing MJCF model in MuJoCo, added a floating base and tabletop contact, and searched over sequences of joint poses. I double-checked the control setup: it did not build a conventional inverse-kinematics pipeline.

On the real hand, those poses were executed through Wuji's SDK with joint-position/PD control. Camera images, joint readings, and IMU feedback were used to check the motion and adjust between attempts. The fingers and thumb provided support as the hand lifted and transferred its weight onto the base. Coordinating those contacts was central to getting it upright.

After the motors switch off

The hand staying upright with its motors disabled is worth a closer look. The final pose seems to make good use of its support contacts, although joint friction may also help. I would try the same task from several starting poses and on a different tabletop. That would help separate a repeatable contact strategy from a sequence that happens to work in this setup.

Interpreting the result

I distinguish the agent's search over poses from the controller that executes them. The observed success belongs to this combined system: the supplied hand model, contact simulation, feedback, and joint controller all contribute. This demonstration does not isolate which component was necessary, or establish that Astra inferred the hand's dynamics accurately.

The Ultra and Max runs also differ in their search histories, so I would not treat their completion times as a controlled model comparison. Repeated trials with matched initial conditions and time budgets would let me report success rates, physical attempts, and failure modes alongside elapsed time.

My posts on X: demonstration · simulation and control details.

@misc{dou2026handselfrighting,
  author = {Dou, Zhiyang},
  title = {Robotic hand self-righting with visual feedback},
  year = {2026},
  month = sep,
  howpublished = {Research Notes},
  url = {https://frank-zy-dou.github.io/blog/wuji2-hand-stands-up/}
}