SWhat changed
The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses.
Scientific scope
The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses. Simulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. Simulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.
Contributors
- The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.6. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction