The adaptive controller ranked highest in the test
The clearest result was the difference in overall task success. Across 180 online trials, the proposed adaptive-stiffness method recorded an average success rate of 95%, compared with 83% for a fixed-stiffness ablation and 69% for a deterministic FACTR baseline. A trial counted as successful when the requested displacement was reached in the instructed direction without a torque-limit shutdown or unintended movement in another direction. The experiment used 15 trials for each of four directions under each of the three conditions.
The gap was not identical in every direction. The proposed method succeeded in 14 of 15 forward trials, 15 of 15 backward trials, 14 of 15 left trials and 14 of 15 right trials. The fixed-stiffness version recorded 10, 14, 14 and 12 successes, respectively, while FACTR recorded 4, 13, 12 and 12. Forward transport showed the largest numerical separation, with the proposed method's reported success rate at 0.93 versus 0.27 for FACTR.
The controller treated disagreement as a control signal
The controller was built around a stochastic generative policy. It took in RGB images and estimates of external joint torque, sampled several possible latent futures, decoded them into action chunks and used the variation among those chunks to adjust joint stiffness and damping. Stiffness is the amount of resistance the robot presents to movement. The study was built around the expectation that greater variation among sampled actions would reduce stiffness and make physical guidance easier, while lower variation would increase assistance. The comparison used a fixed-stiffness version of the policy and a separately trained deterministic FACTR model.
At each inference event, the system generated 10 action chunks, each 100 action steps long and covering about 3.33 seconds. New inference ran at 7.5 Hz, while the robot control loop ran at 500 Hz. In practice, the controller repeatedly looked ahead over a short sequence while continuing to update the robot's response.
The biggest change appeared near the direction choice
In representative forward trials, the action-deviation signal rose during the reported direction-choice interval, time steps 250 to 300, while stiffness fell. The fixed-stiffness condition completed its trial with sustained external joint-torque estimates. In the representative FACTR trial, the robot stopped at the torque safety limit. These examples show the intended contrast in physical response, but they were representative trials rather than an aggregate statistical analysis.
The offline analysis pointed to the same time window in the generated actions. Across 20 validation sequences, posterior variance stayed close to zero when the model received the ground-truth future action chunks, whereas prior variance was larger and changed over time. Prior variance rose especially around time step 250 in latent dimensions 8, 10 and 11, and generated-action variation was especially visible between steps 250 and 300 in Joints 1, 2, 4 and 7. The timing is an association, not proof of a direction-specific mechanism, because task phase and elapsed time were not independently controlled.
A useful result, with a narrow test
The evidence remains limited. The study used one laboratory transport setup, one 7-DoF robot and one human collaborator, with four instructed directions. The online comparison totaled 180 trials, and the reported success rates were descriptive: no confidence intervals or inferential uncertainty estimates were reported. The result therefore does not answer whether the same pattern would hold with different collaborators, robots or tasks.
The underlying demonstrations comprised 120 recorded trials, with 100 used for training and 20 for offline validation. The paper says no random split or random seed was used. That leaves open how stable the result would be under a different partition or a broader evaluation, so the findings are best read as a task-specific feasibility result rather than a general performance claim.
Paper data and sources
Original title: Generative Action-Chunk Sampling for Adaptive Stiffness Control in Physical Human-Robot Collaboration
Authors: Aoi Otake, Ferdinand Hartmann, Ko Igari, Shingo Murata
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text