Skip to content
Bud's Wiki

WUJI Hand 2 autonomous pen spinning with disturbance rejection via RL + motion-reference tracking

TL;DR

Section titled “TL;DR”

A LinkedIn demo post (reshared by CTO Robotics Media from a WUJI Tech team release) shows the 20-DoF WUJI Hand 2 performing autonomous pen spinning while absorbing external disturbances applied by an operator. The system is described as RL-based with motion-reference tracking, with the policy trained in simulation and transferred to the physical hand. No paper, code, or quantitative results accompany the post — it is a product-demo pointer, not a technical disclosure. Its utility is as a datapoint on the current hardware + recipe baseline for autonomous dexterous pen-spin on commercial anthropomorphic hands, directly adjacent to the Dex-SuperMimic — Reproducing SuperMimic on Dexterous Hand Manipulation Dex-SuperMimic open reimplementation and the Wentao Zhu — GPT-6 Astra autonomously trains an Isaac Lab RL policy for pen spinning on the Sharpa hand over 1.5 days GPT-6-Astra-trained Sharpa pen-spin, which share the same task + hardware category.

Key claims

Section titled “Key claims”
  • WUJI Hand 2 is a 20-DoF anthropomorphic dexterous hand demonstrated performing autonomous pen spinning [post body].
  • The controller uses reinforcement learning combined with motion-reference tracking as the training recipe [post body].
  • The policy is trained in simulation and transferred to the physical robot (sim-to-real) [post body].
  • The demo includes external-disturbance rejection: an operator applies perturbations during the spin and the hand continues the manipulation [post body + video].
  • The post is a reshare by CTO Robotics Media (“informational purposes only, does not own or develop the technology”) attributed to the WUJI Tech team: Yutong Hou, Hanya Huang, Ronia Ren, Chenxi Hao [post attribution].

Method

Section titled “Method”

Short LinkedIn caption plus demo video; no paper or code. The described recipe is sim-trained RL policy + motion-reference tracking + sim-to-real transfer, operating over the WUJI Hand 2’s 20 DoF. No details on reward design, the RL algorithm, the reference motion source, domain randomization, observation space (vision vs proprioception vs tactile), or the controller frequency. The disturbance-rejection demo is unquantified — single-take video, no success-rate under perturbation, no comparison to baselines, no description of what disturbances were tested beyond a hand pushing the pen mid-spin.

Results

Section titled “Results”

Qualitative demo only. No reported success rate, no comparison to a hand-tuned baseline, no sim-to-real gap measurement, no indication whether the shown trajectory is cherry-picked or representative. The demo is a company marketing surface rather than a benchmark.

Why it’s interesting

Section titled “Why it’s interesting”

Three places this fits onto the wiki’s existing dexterous-manipulation map.

  1. Hardware substrate. WUJI Hand 1 and WUJI Hand 2 are already the target embodiment in Dex-SuperMimic — Reproducing SuperMimic on Dexterous Hand Manipulation Dex-SuperMimic (which ships a hand2-pen-spin config at 50 Hz using motion-reference + residual control) and feature as a transfer embodiment in AdaDex: Learning Adaptable Dexterous Skills via Hybrid Dexterous Mimicking AdaDex (alongside ArtiMANO/XHand/Sharpa) and in the The Will to Act — Natura-1 Nature Will dual-arm Wuji benchmark against π0.5 and GPT-6-Astra. This LinkedIn demo is the vendor’s own in-house demo of the same hand + task that the Dex-SuperMimic community reimplementation targets — datapoint on what the hardware vendor’s internal recipe looks like next to the community reproduction. Both frame the task identically (motion-reference + residual control over sim-trained RL), reinforcing that this is the current default recipe shape for anthropomorphic in-hand pen-spin.

  2. Pen-spin as a canonical benchmark. Wentao Zhu — GPT-6 Astra autonomously trains an Isaac Lab RL policy for pen spinning on the Sharpa hand over 1.5 days shows the same task on the Sharpa 22-DoF hand being autonomously trained end-to-end by GPT-6 Astra over 1.5 days inside Isaac Lab. The WUJI demo is the hand-engineered counterpart: trained by the vendor’s own research team rather than by an LLM-authored RL harness, same task, slightly different hand (20-DoF vs 22-DoF). Together with Dex-SuperMimic’s hand2-pen-spin checkpoint, pen-spinning on commercial dexterous hands is now a reproducible-across-labs probe that wasn’t the case a year ago.

  3. Disturbance rejection as the differentiator. None of the three sibling pen-spin references (Dex-SuperMimic checkpoint, Astra-trained Sharpa policy, AdaDex’s generalization environment) report explicit reactive disturbance tests. The WUJI post’s framing — “keep spinning while being disturbed” — is a qualitative claim about closed-loop reactivity under external perturbation that the referenced baselines don’t make. That claim is unverified (single video, no quantitative test), but it names a specific capability axis the vendor is positioning on that the open-reimplementation and vendor-neutral RL papers on this page have not yet staked out.

Caveats: this is a marketing surface, not a technical disclosure. No code, no paper, no sim-to-real metric, no description of what the “disturbance” actually was. Treat as a hardware-vendor demo pointer rather than methodological evidence.

See also

Section titled “See also”