A robot hand can learn how to reach, grasp, lift and turn an object, then lose those abilities while being trained for the next job. That is the central warning in ADEPT, a new reinforcement-learning study from NVIDIA and the University of Michigan that treats dexterity less like a collection of isolated tricks and more like a foundation that must survive specialization.
The research, submitted as a preprint on August 19, tested a staged training system on two arm-and-hand platforms. One has 23 controllable degrees of freedom and receives images from two cameras. The other has 29 degrees of freedom, two cameras and five fingertip tactile sensors. Policies were trained in simulation and transferred to the physical robots without additional real-world fine-tuning.
A reusable foundation for robot hands
ADEPT begins with a generic object-reposing task. The robot practices on 16 simple shapes at varied scales, learning a repertoire that includes reaching, grasping, lifting, reorienting and transporting. Researchers then adapt that policy to precise insertion and dish-placement tasks. A geometric controller sits between the learned policy and the hardware to enforce joint limits and collision constraints while leaving the full hand available to the policy.
The cost argument is important. The paper reports roughly eight billion simulated environment steps for pre-training and about three billion more for a downstream specialist. Training one downstream task from scratch consumed about nine billion steps and usually failed to progress as far. Pre-training is only economical, however, if its cost can be shared across many later tasks.
That reuse is not automatic. When the team applied ordinary reinforcement-learning fine-tuning, performance at the starting curriculum level fell to zero within a small number of updates. The old critic had learned to evaluate the reposing reward, not the new insertion reward. Its unreliable estimates pushed the policy away from useful behavior before the critic could catch up.
The proposed repair has three parts: copy the pretrained behavior into the expanded downstream policy, warm up a new critic while freezing the actor, and then make conservative policy updates. In five-seed ablations, every variant using the higher standard learning rate collapsed before completing the curriculum. At the lower rate, critic warm-up improved the reported training success rate by 17.6 percentage points, while behavior cloning reduced average adaptation time from 35.2 hours to 19.9 hours among successful runs.
Analysis: retention is part of capability
The practical bottleneck is not simply teaching a robot one more grip. It is changing the reward, observations and objective without destroying the useful motion already inside the policy. That makes post-training a capability test in its own right. A specialist that reaches the new target only after forgetting how to recover or reorient may score well on a narrow endpoint while becoming less useful as a general machine.
A better scorecard separates four questions: how much foundational skill survives, how much new behavior appears, whether the controller constrains unsafe motion, and whether the robot can close the loop with real sensors. Reporting only final task success compresses those different engineering problems into one number and can conceal where a deployment will fail.
The comparison with the Functional Manipulation Benchmark sharpens that point. The Berkeley-led benchmark was built around reproducible, multi-stage assembly using 54 objects, three boards and 22,550 recorded trajectories. ADEPT evaluates a narrower selection of that world with a different kind of hand. Its reported five-to-ten-second trials are faster than the benchmark’s fixture-assisted parallel-jaw pipeline, but that is a comparison of systems as well as learning methods, not a universal speed record for manipulation.
Touch changes the failure mode
Real-world results also expose the distance between a compelling demonstration and dependable operation. The camera-only robot completed five of ten star-peg trials, three of ten asymmetric-peg trials and six of ten dish placements. On the second platform, adding fingertip touch raised asymmetric-peg completion from three of ten to eight of ten.
That gap suggests the limiting problem was often state uncertainty rather than a missing action sequence. The vision-only policy sometimes made a good grasp but could not confirm contact, reopened its fingers and dropped the object. Touch allowed the system to know that the grasp had succeeded and carry that certainty into lifting, alignment and insertion.
Google DeepMind’s earlier DemoStart work approached simulated dexterity from another direction, using demonstrations to seed a curriculum before transfer to a three-fingered hand. ADEPT removes demonstrations from its tested tasks and emphasizes preservation during post-training. Together, the projects show that simulation is no longer the entire story. The handoff between broad experience, task-specific learning, control constraints and perception now determines how much simulated competence reaches hardware.
What remains unproven
The study is a fresh preprint, and its physical evaluations use ten trials per condition on two stationary research platforms. Occlusion still disrupted estimates of the asymmetric peg’s orientation, while small contact patches made some grasps unstable. The authors also say broader pre-training may be necessary for tasks with little overlap with reposing.
The next meaningful test is not another polished single-task video. It is whether one pretrained foundation can acquire a growing sequence of specialists across more objects and embodiments while retaining earlier skills, respecting the same safety layer and revealing when perception is uncertain. If that standard holds, dexterity becomes reusable infrastructure. If it does not, every impressive robot hand remains a collection of fragile demonstrations.
Sources: ADEPT research paper and project materials from NVIDIA and the University of Michigan; the Functional Manipulation Benchmark from University of California, Berkeley researchers; Google DeepMind’s DemoStart research overview.
Featured image: Robonaut’s dexterous hand opposite an EMU spacesuit glove. Photo: NASA, Robert Markowitz and Bill Stafford, via Wikimedia Commons. Public domain, United States government work. Center-cropped from the 5,479 × 2,740 original to 4,871 × 2,740 at 16:9 and resized to 2,400 × 1,350 pixels; no generative or substantive alteration.
