PaperlayerView exact revision on arXiv

Source-linked AI summary

FACTR: Force-Attending Curriculum Training for Contact-Rich Policy Learning

Jason Jingzhou Liu, Yulong Li, Kenneth Shaw, Tony Tao, Ruslan Salakhutdinov, Deepak Pathak

arXiv:2502.17432v2cs.ROcs.AIcs.CVcs.LG

TL;DR

Contact-rich robot tasks require force feedback, yet teleoperation and policy learning often underuse it. The paper introduces bilateral force-feedback teleoperation and FACTR, which progressively corrupts visual inputs during training to encourage force use. FACTR improves performance and generalization across contact-rich tasks, while the teleoperation system improves user experience and data collection.

  • Problem

    Force feedback is underused in robot teleoperation and policy learning, limiting reliable learning for contact-rich tasks that require force-based adjustments.

  • Method

    The paper combines a bilateral teleoperation system that relays follower-arm torques to actuated leader arms with FACTR, which decreases visual corruption throughout policy training.

  • Results

    FACTR improves task performance and generalization across box lifting, non-prehensile pivoting, fruit pick-and-place, and dough rolling, while the teleoperation system improves reported completion and usability.

  • Takeaways & Limitations

    Force-attending curriculum training helps policies use collected force data instead of neglecting force input through over-reliance on visual information.

  • Takeaways & Limitations

    The demonstrated scope is limited to four contact-rich tasks: box lifting, non-prehensile pivoting, fruit pick-and-place, and rolling dough.

Abstract

from arXiv · show

Many contact-rich tasks humans perform, such as box pickup or rolling dough, rely on force feedback for reliable execution. However, this force information, which is readily available in most robot arms, is not commonly used in teleoperation and policy learning. Consequently, robot behavior is often limited to quasi-static kinematic tasks that do not require intricate force-feedback. In this paper, we first present a low-cost, intuitive, bilateral teleoperation setup that relays external forces of the follower arm back to the teacher arm, facilitating data collection for complex, contact-rich tasks. We then introduce FACTR, a policy learning method that employs a curriculum which corrupts the visual input with decreasing intensity throughout training. The curriculum prevents our transformer-based policy from over-fitting to the visual input and guides the policy to properly attend to the force modality. We demonstrate that by fully utilizing the force information, our method significantly improves generalization to unseen objects by 43\% compared to baseline approaches without a curriculum. Video results, codebases, and instructions at https://jasonjzliu.com/factr/

I. INTRODUCTION

The paper addresses underused force feedback in contact-rich robot learning by presenting bilateral force-feedback teleoperation and FACTR, a curriculum that shifts policy attention toward force while visual corruption decreases. Across four contact-rich tasks, the approach improves teleoperation usability and policy generalization to varied objects.

  • Motivation: Contact-rich tasks require precise force-based adjustments, but robot learning commonly prioritizes vision and therefore struggles to generalize to novel object appearances and geometries.Force feedback can signal contact and mode changes when visual attributes such as color or texture are irrelevant.
  • Teleoperation: The proposed bilateral teleoperation system actuates leader-arm motors using follower-arm joint torques, adding force feedback to data collection for dynamic contact-rich tasks.The system also provides active gravity compensation and redundancy resolution.
  • Teleoperation: 64.7% higher task completion, 37.4% shorter completion time, and 83.3% better subjective ease of use were achieved across four evaluated contact-rich tasks.These gains were reported for the force-feedback teleoperation experience.
  • Policy Learning: FACTR progressively reduces visual corruption during training so policies learn to use force information before receiving higher-fidelity visual inputs.The curriculum applies Gaussian blurring or downsampling with a scheduler that decreases corruption over time.
  • Results: 40.0% better performance was reported than policies given force inputs but trained without FACTR, across four challenging contact-rich tasks.The evaluated tasks were box lifting, non-prehensile pivoting, fruit pick-and-place, and dough rolling.
  • Prior Systems: The system uses low-cost actuated leader arms and grippers, while related leader-follower systems typically lack force feedback or bilateral information transfer.The design retains low-cost hardware while adding force feedback, gravity compensation, and redundancy handling.

A. Force Feedback

The system relays follower-arm external torques to the leader, while adding gripper feedback, customizable redundancy resolution, gravity compensation, friction compensation, and joint-limit protection for teleoperation.

  • Force Feedback: Follower-arm external joint torques are relayed to the leader so operators can feel the follower’s physical constraints.The feedback uses scaled external torque and damping to reduce leader-arm oscillations during contact.
  • Force Feedback: The servo-based gripper uses its current reading to provide force feedback through a smoothed signal.The system sets the exponential moving-average smoothing factor α to 0.1.
  • Customizable Redundancy Resolution: A null-space projection regulates joints toward a user-defined resting posture without imposing additional end-effector wrenches.This addresses reachability problems caused by fixed or mechanical joint regularization in confined spaces.
  • Gravity Compensation: Gravity compensation keeps the leader stationary by computing torques that counteract dynamic forces with recursive Newton-Euler inverse dynamics.The modeled terms include inertia, Coriolis and centrifugal effects, and gravity.
  • Additional Compensation and Controls: Friction compensation, artificial joint-limit potentials, and inter-arm obstacle avoidance further constrain and smooth leader-follower operation.Riemannian Motion Policies are used for dynamic obstacle avoidance between bimanual follower arms.

E. Overall Control Law for the Leader Arm

The leader-arm control combines force feedback, null-space regulation, gravity and friction compensation, and joint-limit protection, while the learned policy maps visual and force tokens to future joint-position targets.

  • Overall Control Law: The combined leader-arm torque is τ = τfeedback + τnull + τgrav + τfriction + τlimit.These terms respectively provide follower-force feedback, null-space regulation, gravity compensation, friction compensation, and joint-limit protection.
  • Customizable Joint Regularization: Customizable resting posture improves reachability in confined spaces, where fixed regularization can lead to collisions.The leader arm lets the user define qrest for the follower’s desired configuration.
  • Policy Model: The policy receives images and external joint torques, then predicts future joint-position targets using behavior cloning.The target sequence spans k future time steps and is trained against expert joint-position targets.
  • Policy Model: Vision and force observations are tokenized, encoded, and processed by an encoder-decoder transformer whose action tokens cross-attend to both modalities.The vision encoder produces vision tokens, while an MLP force encoder produces a force token.
  • Policy Model: The policy projects decoder outputs into joint-position action space for the follower arm.The action representation has dimension d_a, and the behavior-cloning objective compares predicted and expert future targets.

B. Force-Attending Curriculum

FACTR addresses force underuse by corrupting visual inputs during early training and gradually restoring visual fidelity, encouraging the policy to learn force-dependent behavior first.

  • Curriculum Design: FACTR applies Gaussian blurring or downsampling to visual inputs with a scale that decreases throughout training.The operators can act in pixel space or on visual latent tokens.
  • Curriculum Design: The training algorithm samples demonstrations, updates the corruption scale with a scheduler, and applies the selected pixel- or latent-space operator before encoding observations.The curriculum runs over N gradient steps using a pretrained vision encoder, force encoder, and action-chunking transformer.
  • Curriculum Design: At very large corruption scales, visual inputs become approximately indistinguishable, so early gradients emphasize force information and force-encoder updates.This is the stated intuition for learning force-dependent distinctions before detailed visual information returns.

C. Curriculum Operators

FACTR uses Gaussian blur or downsampling in pixel or latent space, gradually reducing the corruption scale so the policy first emphasizes force and later fuses force with detailed vision.

  • Curriculum Operators: FACTR considers Gaussian blur and downsampling as curriculum operators for degrading visual information.The operators can be applied to images or visual latent tokens.
  • Gaussian Blur: Gaussian blur applies a scale-dependent kernel through convolution in pixel space or latent space.The pixel-space version uses a 2D kernel, while the latent-space version uses a 1D kernel.
  • Downsampling: Downsampling uses max pooling followed by nearest interpolation in pixel space, with an analogous one-dimensional latent-space operator.The latent-space version uses MaxPool1D.
  • Curriculum Schedule: Reducing σ_n makes the policy focus first on force tokens and then incorporate visual information later in training.The stated outcome is more robust force-vision fusion and less overfitting to vision.
  • Theoretical Analysis: A Neural Tangent Kernel analysis studies Gaussian blur in a simplified two-layer model, with the intuition intended to extend to vision transformers.The formal analysis is presented in the appendix.

D. Curriculum Schedulers

The evaluation uses four contact-rich tasks with varied objects and visual appearances, requiring policies to perceive and respond to force feedback during manipulation.

  • Tasks: The study evaluates policies on four contact-rich tasks: box lifting, non-prehensile pivoting, fruit pick-and-place, and rolling dough.
  • Tasks: Box lifting requires two arms to lift and balance a box in the air for at least two seconds using visual observations and external joint torque.
  • Tasks: Non-prehensile pivoting rotates an item by 90° against a fixture corner until it stands stably, using visual observations and external joint torque.
  • Tasks: Fruit pick-and-place grasps a soft, delicate fruit and places it in a bowl using a wrist camera and gripper torque.
  • Tasks: Rolling dough continuously shapes the dough into a cylinder for at least 8 seconds using visual observations and arm torque.

B. Teleoperation Evaluation

The evaluation compares force-feedback teleoperation and FACTR-based policies with baselines across contact-rich tasks, showing benefits for teleoperation usability and object generalization.

  • Teleoperation: The proposed teleoperation system is compared with an un-actuated leader-follower baseline with mechanical joint regulation.
  • Teleoperation: 64.7% higher task completion rate, 37.4% reduced completion time, and 83.3% improvement in subjective ease of use were observed with the proposed system.
  • Teleoperation: Without force feedback, the baseline often loses contact during continuous-contact tasks because operators cannot perceive the environment’s geometric constraints.
  • Teleoperation: Force feedback prevents operators from moving the leader arms too far from the follower arms during environmental contact.
  • Policy evaluation: On test objects, vision-only policies achieve 21.3% average success, force-input policies without a curriculum achieve 61.2%, and FACTR achieves 87.5%.
  • Policy evaluation: During box lifting, FACTR attends more to force after contact, whereas policies without the curriculum often fail to lift or balance novel boxes.
  • Policy evaluation: For test objects, vision-only recovery success falls from 31.7% on the first attempt to 13.3% on the second, while force-attending policies maintain similar rates.

D. Ablations on Curriculum

The ablations examine curriculum operators and schedulers, finding that decaying visual smoothing outperforms fixed smoothing while performance remains relatively robust across parameter choices.

  • Ablation setup: The ablation study evaluates pixel-space and latent-space curricula, Gaussian blur and downsampling, and four scheduler types on pivoting test objects.
  • Fixed-scale comparison: A curriculum with decaying smoothing performs better across all tasks than fixed smoothing.
  • Fixed-scale comparison: Fixed smoothing can prevent policies from extracting the necessary details from fully unblurred vision, despite reducing visual overfitting.
  • Parameter comparisons: No scheduler, operator, or curriculum-space choice shows a uniform advantage or disadvantage, suggesting relative robustness to curriculum parameters.
  • Conclusion: FACTR uses decreasing visual blur to encourage force use early in training and improve generalization across unseen object appearances and geometries.
  • Limitations: The method depends on external joint torque sensors, and its curriculum hyperparameters can require task-dependent tuning.

APPENDIX

The appendix uses Neural Tangent Kernel analysis to explain how Gaussian blur changes visual similarity and why decreasing blur can shift learning from force-relevant global structure toward finer visual features.

  • NTK analysis: FACTR analyzes Gaussian blur with NTK as a theoretical framework, using a simple two-layer model while suggesting the intuition extends to vision transformers.The analysis treats Gaussian blur as an example curriculum operator.
  • NTK preliminaries: The NTK measures similarity through parameter-gradient inner products and becomes deterministic in the infinite-width limit under Gaussian initialization.Its kernel matrix is positive semidefinite, with entries Kij = k(xi, xj).
  • Effect of Gaussian blur: As blur increases, Gaussian convolution suppresses high-frequency information, makes inputs more similar, and drives the NTK toward an approximately constant all-ones kernel.In the limit σ →∞, blurred inputs converge to the same vector and lose pairwise discrimination.
  • Curriculum interpretation: FACTR decreases blur over training, exposing the model first to smoother visual patterns and later to fine-grained visual features while encouraging early reliance on force information.The appendix states that early updates focus on force information and update the force encoder to differentiate inputs.
  • NTK dynamics: Training residual components parallel to an eigenvector decay exponentially at the corresponding eigenvalue rate, while perpendicular components remain unchanged.The residual decomposition motivates analyzing which input distinctions the kernel preserves.

VIII. COST ANALYSIS OF OUR TELEOPERATION SYSTEM WITH FORCE FEEDBACK

The teleoperation system is designed as a low-cost force-feedback platform, with documented component costs and additional control laws for reliable operation.

  • Cost: The bill of materials reports a total cost of around $1229.95 for one leader-arm teleoperation device with force feedback.The pricing is described as accurate as of the paper release.
  • Control additions: The system adds friction compensation and joint-limit avoidance as control laws for force-feedback teleoperation.These controls address motor resistance and prevent the leader arm from exceeding follower-arm joint limits.

A. Friction Compensation

The teleoperation controller compensates static, Coulomb, and viscous friction, while adding a repulsive potential to maintain joint-limit safety.

  • Friction model: Joint friction reduces responsiveness and increases operator effort, so the system uses a dynamic model covering static, Coulomb, and viscous friction.The compensation is intended to preserve precise motion control and reduce fatigue during prolonged operation.
  • Static friction: Static friction is mitigated with a high-frequency oscillatory control signal that generates micro-vibrations and smooths transitions from rest to motion.The control-loop frequency is set to 500 Hz.
  • Kinetic friction: Kinetic friction compensation combines Coulomb and viscous friction terms using their respective coefficients.The total friction compensation is the sum of static and kinetic terms.
  • Joint-limit avoidance: An artificial potential-based control law prevents the operator from moving the leader arm beyond the follower arm’s joint limits.The repulsive potential uses a safety margin and scaling factor.

C. Bi-manual Follower Arms Control with Dynamic Collision Avoidance

The system combines dynamic follower-arm control with a transformer behavior-cloning policy, while evaluating generalization across tasks and unseen objects.

  • Follower-arm control: A Riemannian Motion Policy dynamically generates follower joint targets while incorporating real-time collision avoidance.This replaces direct position-target copying from leader to follower arms.
  • Policy architecture: The behavior-cloning policy uses image tokens from a ViT and proprioceptive tokens from a linear layer, then predicts action sequences with an encoder-decoder action-chunking transformer.Excluding current arm joints from proprioception is reported to improve generalization by discouraging near-current-state action prediction.
  • Training: Policy training generally uses 20000-50000 gradient steps and 2-6 hours of wall-clock time on an RTX4090.The architecture and training hyperparameters are listed in Table IV.
  • Alternative methods: Adaptive normalization caused training instability and overfitting, while low-level visual augmentation underperformed FACTR on test objects and higher levels failed across training and testing.The curriculum is described as avoiding the tradeoff between using vision and using force.
  • Evaluation: The evaluation covers box lift, non-prehensile pivot, fruit pick-place, and rolling dough, with comparisons reported in Tables VII-X.Table VI evaluates unseen objects across three tasks, and the testing set was doubled for three tasks.
Loading 2502.17432v2…