For the proposed condition, the reported average success rate was 0.95, compared with 0.83 for a fixed-stiffness version of the same stochastic policy and 0.69 for a separately trained deterministic FACTR controller. The system was designed to change how much help the robot gives a person when more than one future movement remains plausible. In the laboratory object-transport task, it used the spread among its own sampled future movements to adjust joint stiffness and damping. In ordinary terms, stiffness describes how strongly the arm resists movement, while damping describes how much it restrains motion as movement unfolds. Those figures describe the tested setup; they do not establish performance beyond the task and directions examined.
The experiment paired one human collaborator with a robot manipulating an object toward four target directions. It was a single collaborative transport setup, not a broad assessment across people or tasks.
The signal came from a spread of predictions
The system took RGB images and external joint-torque estimates as input. Inference ran at 7.5 Hz, roughly every 133 milliseconds. At each pass it generated 10 action chunks, each containing 100 action steps, from an observation-conditioned prior. All 10 samples contributed to the estimate of empirical variation, while one chunk was selected as the control candidate. The variation was then used to adapt joint stiffness and damping.
The model's prior variance tended to increase around time step 250, especially in latent dimensions 8, 10 and 11. The sampled action trajectories showed pronounced variation between steps 250 and 300, particularly in Joints 1, 2, 4 and 7. The reported co-occurrence indicated that changes in the latent distribution were reflected in action predictions, although task phase and elapsed time were not independently controlled.
The training material came from the same narrow experiment. The sole collaborator was the paper's first author, who provided 120 demonstration trials. Researchers manually stratified those data into 100 training trials and 20 offline-validation trials, without a random split or random seed. Ten offline-validation trials separated from training were used to calibrate the minimum and maximum action-deviation thresholds, which ended at 0.061 and 0.216.
The hardware setup used a 7-DoF Franka Research 3 arm, a 640 by 480 RGB camera, manufacturer-API estimates of external joint torque and a joint controller running at 500 Hz. It did not use direct Cartesian contact-force measurements. The force-related input in this experiment was therefore a joint-torque estimate, not a direct reading of contact force.
The clearest gap came in forward movement
Online evaluation compared three conditions: the proposed adaptive-stiffness system, the same stochastic policy with Medium stiffness and damping fixed, and separately trained deterministic FACTR with Medium settings fixed. The models were not retrained for the comparison. Each condition had 15 trials in each direction, and the 180 online trials followed a repeating Right, Forward, Left, Backward cycle.
Across the proposed condition's direction-specific trials, the reported average success rate was 0.95. The fixed-stiffness ablation averaged 0.83, and deterministic FACTR averaged 0.69. The direction counts were 14 of 15 forward, 15 of 15 backward, 14 of 15 left and 14 of 15 right for the proposed system. The ablation recorded 10, 14, 14 and 12 successes in that order, while FACTR recorded 4, 13, 12 and 12.
The gap was especially clear in forward transport. The proposed method's success rate was 0.93, compared with 0.27 for FACTR. In a representative trace from the proposed method, action deviation increased and stiffness decreased around steps 250 to 300. The displayed FACTR trial stopped at the torque safety limit. Because the trace was representative rather than an aggregate analysis, it illustrates the reported response without showing how often that pattern occurred.
Rightward results were closer. The proposed method succeeded in 14 of 15 trials, compared with 12 of 15 for both the ablation and FACTR. The report says some comparator failures transitioned into backward motion after rightward transport. In proposed trials, return motion was manually corrected with reduced stiffness. The failure inspection was qualitative, and the explanation based on visual similarity was not independently tested.
A result with a narrow frame
Several design choices narrow what the numbers can tell us. The action-deviation thresholds were empirically calibrated and may need retuning for other tasks or hardware. Only the first author served as collaborator, and the evaluation directions followed a fixed repeating cycle. The study reported descriptive rates without confidence intervals or inferential tests. It therefore does not establish generalization beyond the tested transport task and four directions.
The variation signal also has a narrower meaning than a probability of human intention. It was not calibrated as a human-intention probability, and it did not identify epistemic uncertainty under out-of-distribution observations. The evaluation did not establish generalization beyond the tested transport task and its directional outcomes.
The comparison also leaves a specific question about what drove the difference. The proposed-versus-ablation test gave the proposed system access to a wider Low-to-High stiffness range, while the FACTR comparison included policy-architecture differences. The results therefore describe the behavior of the tested combinations, not an isolated causal effect of adaptive stiffness.
The paper is a preprint. Its front matter says the work was submitted to IEEE for possible publication and that its human-subject procedures were exempt from review-board approval. It reports support from JST PRESTO, JST Moonshot R&D, JSPS KAKENHI and the Kayamori Foundation of Informational Science Advancement.
The findings therefore describe a task-specific control behavior in one laboratory transport task with one collaborator and four directions. They do not establish a general result about human-robot collaboration or a calibrated measure of human intent.
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