Generative Action-Chunk Sampling for Adaptive Stiffness Control in Physical Human-Robot Collaboration

Aoi Otake; Ferdinand Hartmann; Ko Igari; Shingo Murata

Source-linked AI summary

Generative Action-Chunk Sampling for Adaptive Stiffness Control in Physical Human-Robot Collaboration

Aoi Otake, Ferdinand Hartmann, Ko Igari, Shingo Murata

arXiv:2608.25284v1Aug 26, 2026cs.ROcs.AIcs.LG

TL;DR

Physical human–robot collaboration requires assistance when intention is clear but compliance when several future motions remain plausible. The paper samples action chunks from an observation-conditioned multimodal prior and uses their variation to adapt stiffness and damping. In four-direction collaborative transport, it achieved a 0.95 average success rate versus 0.83 with fixed stiffness and 0.69 for deterministic FACTR.

Problem
The controller must balance useful assistance with compliance because human intentions can change and several future motions may remain plausible.
Method
A multimodal generative policy samples multiple future action chunks from an observation-conditioned prior, using their variation to continuously adapt stiffness and damping.
Results
0.95 average success rate was achieved in four-direction collaborative transport, compared with 0.83 for fixed stiffness and 0.69 for deterministic FACTR.
Takeaways & Limitations
Sampled-action variation can serve as an online control signal for balancing action tracking and physical compliance in pHRI.
Takeaways & Limitations
The evaluation does not establish generalization beyond the tested collaborative transport task and directional outcomes.

Abstract

from arXiv · show

Physical human-robot collaboration requires a robot to provide assistance when human intention is clear while remaining compliant when several future motions are plausible. We present an adaptive stiffness framework based on generative action-chunk sampling. Conditioned on an RGB image and external joint-torque estimates, the policy samples multiple future action chunks from an observation-conditioned prior. Variation among the sampled action chunks is used to continuously adapt joint stiffness and damping. Greater variation makes the robot more compliant to facilitate human guidance, whereas lower variation provides firmer assistance. In a real-world collaborative transport task with four possible directions, the proposed method achieved an average success rate of 0.95, compared with 0.83 for a fixed-stiffness ablation and 0.69 for a deterministic baseline. Near direction determination, variation among the sampled action chunks increased and the controller accordingly reduced stiffness. These results suggest that variation among actions sampled by a generative policy can serve as an online control signal for balancing assistance and compliance in physical human-robot interaction.

I. Introduction

The paper addresses the need to balance robotic assistance and compliance when human motion is changing or ambiguous. It proposes generative action-chunk sampling whose variation drives adaptive stiffness and damping in collaborative transport.

High stiffness can resist unanticipated human motion, while uniformly low stiffness limits physical support during collaboration.
Sampled-action variation continuously adjusts stiffness and damping: greater variation lowers stiffness for guidance, whereas lower variation increases stiffness for assistance.
The framework extends multimodal force-aware action chunking with ACT-style CVAE training and an observation-conditioned prior.
The proposed policy combines visual and external joint-torque observations to sample multiple plausible future action chunks.

B. Modeling Intention Ambiguity in Human–Robot Interaction

The framework represents ambiguity through distributions over future actions and connects multimodal generative prediction to variable impedance control. Its sampled action variation is used as an online signal for yielding or assisting during interaction.

Large variation among sampled action chunks indicates that multiple future motions remain plausible, whereas small variation indicates similar supported futures.
Lower stiffness can reduce resistance to human corrections, while higher stiffness can improve tracking and assistance after the desired motion is established.
The same generative policy supplies both the target action chunk and the variation-based stiffness and damping control signal.
The policy encodes RGB images and external joint-torque estimates into a multimodal Transformer representation for action-chunk prediction.
ACT-style CVAE training uses a learned observation-conditioned prior so multiple latent variables and future action chunks can be sampled during inference.

C. Training Procedure

Training optimizes the policy using a combined reconstruction and KL-divergence objective.

The network parameters are optimized by minimizing a linear combination of reconstruction loss and Kullback–Leibler divergence loss.

1) Loss Function:

The loss function combines action-chunk reconstruction with KL alignment of the observation-conditioned prior and approximate posterior.

1) Loss Function:: The reconstruction loss is the mean absolute error between predicted and ground-truth action chunks in normalized action space.
1) Loss Function:: The KL divergence measures statistical distance between the approximate posterior q_φ and prior p_ψ.
1) Loss Function:: KL minimization aligns the observation-only prior with the ground-truth-conditioned posterior, enabling appropriate latent sampling when ground-truth data is unavailable.
1) Loss Function:: Curriculum learning is adopted to prevent overfitting to visual inputs and integrate force data effectively.

2) Curriculum Learning:

The controller quantifies variation across multiple sampled action chunks, emphasizes later predictions, and smooths the resulting signal before mapping it to stiffness adaptation.

2) Curriculum Learning:: Multiple sampled action chunks provide an action-deviation signal that acts as a proxy for uncertainty when alternative futures remain plausible.Similar samples indicate compatible futures, whereas varying samples indicate multiple plausible motions.
2) Curriculum Learning:: The controller computes empirical action variance at each chunk offset and converts it with a square-root transformation into action-deviation magnitude.The calculation uses the mean normalized action, Bessel’s correction, and ε = 10^-8 for numerical stability.
2) Curriculum Learning:: An exponential moving average smooths the action-deviation signal before normalization for control.The implementation uses γ = 0.1, retaining 90% of the historical state, with empirical bounds determined from offline validation data.
2) Curriculum Learning:: The normalized action-deviation signal is used to derive a stiffness blending coefficient between high- and low-stiffness settings.The supplied passage introduces this mapping but does not provide the full coefficient equation.

3) Calculation of Adaptive Stiffness Parameters:

The controller limits changes in its stiffness blending coefficient and interpolates joint stiffness and damping between predefined high and low profiles.

3) Calculation of Adaptive Stiffness Parameters:: A slew-rate limiter converts the desired blending coefficient into the actual control coefficient at each low-level control step.The implementation uses Δα_max = 0.005 per step and initializes α_0 = 1.0.
3) Calculation of Adaptive Stiffness Parameters:: The stiffness matrix K_r and damping matrix D_r are linearly interpolated between predefined high-stiffness and low-stiffness parameters.The desired coefficient assigns α̂_t = 1 to high stiffness and α̂_t = 0 to low stiffness.

IV. Experiments

The experiments use a four-direction collaborative transport task in which a human and robot lift an object together, determine a direction through human force, and transport it cooperatively.

IV. Experiments: The task requires the human and robot to jointly move an object toward one of four target directions: forward, backward, left, or right.The direction alternatives are introduced during the transport task after collaborative lifting.
IV. Experiments: The robot and human then cooperatively transport the object in the determined direction.The task sequence is approach and lifting, direction determination and guidance, followed by transport.
IV. Experiments: The robot approaches the box, and the human and robot sandwich it from opposite sides to lift it vertically.The lightweight, sponge-based object can be held through friction without finger grasping.
IV. Experiments: During direction determination, the human applies force to indicate one of the four intended transport directions.This phase follows lifting and precedes cooperative transport in the selected direction.

B. Hardware Setup

The system uses a 7-DoF robot with RGB and external joint-torque observations, trains on manually stratified demonstrations, and infers multiple action chunks for adaptive impedance control.

B. Hardware Setup: The hardware setup uses a Franka Research 3 robot arm, RGB images, and external joint-torque estimates from the manufacturer’s API.The force input does not comprise direct Cartesian contact-force measurements.
B. Hardware Setup: The observations are recorded at 30 Hz, with 640 × 480 RGB images resized to 224 × 224 before model input.The same human collaborator collected demonstrations and participated in real-world evaluation.
C. Data Collection: 120 demonstration trials are manually split into 100 training trials and 20 offline validation trials without a random split or random seed.Trials combine Low and High stiffness settings with four directional movement variations.
C. Data Collection: The Low and High stiffness and damping profiles are empirically tuned using preliminary physical-followability evaluations under the present conditions.The settings aim to provide responsiveness while mitigating excessive force conflicts.
2) Inference and Control Settings:: Each inference event generates M = 10 action chunks of 100 action steps, spanning approximately 3.33 s, at an inference rate of 7.5 Hz.All samples compute empirical action variance, while one sampled chunk serves as the control candidate.
2) Inference and Control Settings:: Temporally overlapping predictions are combined into the target joint command using exponentially weighted averaging, with older predictions receiving greater weight.At N_buf = 8, the oldest prediction contributes approximately 25.6%, while the newest contributes about 4.4%.

E. Baselines and Evaluation Metrics

The evaluation compares stochastic adaptive control with fixed-stiffness and deterministic conditions, while latent-variance analysis examines how prior and posterior uncertainty evolve over time. Prior variance increased near direction determination, whereas posterior variance remained concentrated.

Baselines and evaluation setup: Three conditions were compared: the proposed stochastic policy with adaptive stiffness, the same policy with fixed stiffness, and deterministic FACTR with fixed stiffness.Neither model was retrained or fine-tuned during evaluation.
Baselines and evaluation setup: 180 online trials evaluated success across four transport directions, using direction-specific displacement thresholds.Each condition included 15 trials in each direction.
Latent-space variance: The plotted variance parameters were obtained by exponentiating network-output log variances across 20 offline validation sequences.Thin lines represent individual sequences, while thick lines represent their mean at each time step.
Latent-space variance: Posterior variance remained close to zero throughout the task, indicating concentration when the ground-truth future action chunk was provided.Prior variance was larger and varied over time because it used only current observations.
Latent-space variance: Prior variance tended to increase around time step 250, particularly in latent dimensions 8, 10, and 11, during direction determination.The association is not isolated from task phase or elapsed time.

B. Analysis of Sampled Action Chunks

Sampled action chunks varied most near the behavioral bifurcation, consistent with changing prior variance, while the real-world evaluation compared task outcomes across transport directions and methods.

Analysis of Sampled Action Chunks: Pronounced variation appeared in generated joint trajectories between steps 250 and 300, especially for Joints 1, 2, 4, and 7.These differences occurred around the behavioral bifurcation.
Analysis of Sampled Action Chunks: Increased prior variance and greater trajectory variation occurred together, indicating that latent-distribution changes were reflected in action predictions.This observation links the sampled latent variables to generated joint trajectories.
Real-World Evaluation: 0.95 was the proposed method’s overall average success rate, compared with 0.83 for fixed stiffness and 0.69 for deterministic FACTR.The comparison used the proposed policy, its fixed-stiffness ablation, and deterministic FACTR.
Interpretation and caveat: The FACTR comparison does not isolate conditional-prior sampling because it also includes other architectural differences.FACTR recorded fewer successes than the proposed method in all four directions.
Real-World Evaluation: 14/15 successes were achieved by the proposed method in the Right direction, versus 12/15 for both the ablation and FACTR baseline.All methods had relatively high success rates in that direction.
Real-World Evaluation: 0.93 was the proposed method’s Forward success rate, compared with 0.27 for the FACTR baseline.The largest numerical difference among methods occurred in the Forward direction.

1) Analysis of Forward Transport:

Representative trials showed increased action variation and reduced stiffness near direction determination, while the proposed method achieved higher average success than fixed-stiffness and deterministic baselines. The findings are qualified by task-specific evaluation and empirically selected scaling thresholds.

Adaptive response: Action-deviation variation increased and stiffness decreased around the bifurcation interval at time steps 250–300.This pattern was observed in the representative proposed-method trial.
Results: Rightward transport achieved 14/15 successes with the proposed method and 12/15 with both the ablation and FACTR.Some ablation and FACTR failures transitioned into backward motion after rightward transport.
Failure analysis: Visual similarity near direction determination and task completion likely contributed to unintended return motion in some ablation and FACTR trials.The proposed method’s reduced stiffness allowed participant correction when a similar misinterpretation occurred.
Results: 0.95 average success rate was achieved by the proposed method, compared with 0.83 for the fixed-stiffness ablation and 0.69 for deterministic FACTR.The comparison was reported for the evaluated collaborative transport task.
Limitations: The action-deviation signal is not a calibrated probability of human intention and does not identify epistemic uncertainty under out-of-distribution observations.The evaluated collaborative transport setting also does not establish generalization beyond the tested task and directional outcomes.
Limitations: The scaling thresholds σmin and σmax were selected empirically and may require retuning for different tasks or robot hardware.Future work proposes automated calibration and broader collaborative evaluation.