An arXiv preprint posted Sept. 30 reports that scheduling action noise during training helped a robot hand place objects at target positions more often, in both simulation and a physical test. The team tested three learning methods on a RealMan RM75 arm with an Inspire/RH56 hand; the 360 physical trials covered three objects, with 20 attempts for each baseline and scheduled-policy condition per method–object pair.
Action noise adds variation to a policy’s commands as it learns which finger movements and contacts can move an object. DexPolicy starts with broader exploration and gradually narrows that noise according to training steps. It does not change the noise separately during approach, grasp and lift. In each paired comparison, the researchers kept the loss, architecture, reward and optimizer settings fixed, then ran both policies using their learned action means without sampling noise.
On the physical robot, average Target success rose from 8.3% to 43.3% for PPO, 10.0% to 63.3% for GRPO, and 25.0% to 85.0% for FPO. Target required the hand to make the required contacts and leave the object within 3 cm of its target. The separate Lift measure counted an object held more than 5 cm high, so a successful grasp did not automatically count as successful placement. Each method’s rates average 60 trials across a tomato can, sugar box and mustard bottle.
In simulation, mean Target success across five objects and three training seeds also increased: from 32.0% to 35.7% for PPO, 14.1% to 45.4% for GRPO, and 49.4% to 68.1% for FPO. These are comparisons within each method; the study did not run an equal-budget contest between the three algorithms.
The hardware result comes from one trained policy for each baseline and scheduled object–method condition, so its 20 attempts per condition do not show how the result varies across independent training runs. The work also tests only three physical objects in one trajectory-guided setup. The authors’ code release contains the method core, but not the simulator, object assets, weights, robot drivers or real-robot data needed to reproduce the reported tests.
Sources
- DexPolicy preprint (arXiv:2610.00360)
- Authors’ method-code repository
- Image: Robot hand at VivaTech 2019, by Jérémy Barande / École polytechnique, CC BY-SA 2.0. Representative image, not the hardware in the study.
.jpg){kind=link}