Venue: ICLR 2026 · Authors: Kehlani Fay, Darin Anthony Djapri, Anya Zorin, James Clinton, Ali El Lahib, Hao Su, Michael T. Tolley, Sha Yi, Xiaolong Wang (UC San Diego / UC Santa Barbara) · arXiv:2512.03743 · Category: Dexterous Manipulation · Trend tag: Morphology-policy co-design

flowchart LR
  Space[Morphology search space<br/>joints, fingers, palm] --> CC[Morphology-conditioned<br/>cross-embodied control]
  CC --> Eval[Scalable eval across<br/>wide design space]
  Eval --> Sel[Select task-specific<br/>hand morphology + policy]
  Sel --> Fab[Fabricate with<br/>accessible components]
  Fab --> Deploy[Real deployment<br/>< 24h end-to-end]

Dexterous-hand research usually fixes the hardware (a given hand) and only optimizes the controller. But the right morphology is task-dependent, and hand-designing morphologies is slow. The open question: can we jointly optimize hand morphology and control so a hand is purpose-built for a task — and do it fast enough to be practical?

House of Dextra is a co-design framework that jointly learns a task-specific hand morphology and a complementary dexterous control policy. Key pieces:

• Expansive morphology search space spanning joint, finger, and palm generation — not just parameter tweaks but structural variation.
• Morphology-conditioned cross-embodied control: a single controller conditioned on the morphology lets the system evaluate many candidate designs without training a separate policy per design, making search over the large space scalable.
• Real-world fabrication with accessible components, closing the loop from design to physical hand.

The headline systems result: an end-to-end pipeline that designs, trains, fabricates, and deploys a new robotic hand in under 24 hours. The framework is stated to be open-sourced.

• Evaluated across multiple dexterous tasks, including in-hand rotation, in both simulation and real-world deployment.
• Demonstrates that co-designed, task-specific morphologies paired with morphology-conditioned policies transfer to fabricated hardware.

(Specific per-task quantitative metrics are reported in the full paper.)

Pushes dexterous manipulation from "design the controller for a fixed hand" toward "design the hand and the controller together," with a turnaround fast enough (<24h) to be a real engineering tool. The morphology-conditioned cross-embodied controller is the enabling trick — it amortizes policy learning across the whole design space, which is what makes large morphology search tractable.

• arXiv:2512.03743
• OpenReview

• ICLR 2026 Survey
• Dexterous Manipulation review
• DexNDM
• VLMgineer

← Back to ICLR-2026