Adaptive Motion Planning via Contact-Based Intent Inference for Human-Robot Collaboration

Basic Information

• Paper ID: 2510.08811
• Title: Adaptive Motion Planning via Contact-Based Intent Inference for Human-Robot Collaboration
• Authors: Jiurun Song (Texas A&M University), Xiao Liang (Texas A&M University), Minghui Zheng (Texas A&M University)
• Classification: cs.RO (Robotics)
• Publication Date: October 9, 2024 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.08811

Abstract

Human-robot collaboration (HRC) requires robots to adapt to human intent to ensure safe and efficient cooperation in shared spaces. While large language models (LLMs) provide high-level reasoning capabilities for inferring human intent, their application in reliable motion planning for HRC remains challenging. Physical human-robot interaction (pHRI) is intuitive but typically relies on continuous kinesthetic guidance, imposing a burden on operators. To address these challenges, this paper introduces an adaptive motion planning framework based on contact information that directly infers human intent from physical contact for online motion correction in HRC.

Research Background and Motivation

Core Problem

In human-robot collaboration, robots must adapt to human intent in real-time to ensure safe and efficient cooperation, particularly in uncertain and dynamic environments. When the robot's planned path conflicts with human intent, timely adjustments are necessary.

Problem Significance

1. Safety Requirements: In shared workspaces, robots must respond to human safety corrections
2. Efficiency Demands: Reduce the burden of continuous physical guidance by humans and improve collaboration efficiency
3. Practical Considerations: Existing methods either rely on expensive sensors or require continuous human intervention

Limitations of Existing Approaches

1. Large Language Models: Suitable for high-level task planning but suffer from latency and incomplete descriptions in continuous low-level motion adaptation
2. Kinesthetic Guidance: Requires sustained physical effort, imposing burden on operators
3. Tactile Sensors: Complex deployment, high cost, and challenges in surface conformability and impact resistance

Research Motivation

Exploit the natural encoding of human intent in the magnitude and direction of contact forces, conveying correction intent through brief contact rather than continuous guidance.

Core Contributions

1. Optimization-Driven Force Estimation Method: Infer contact forces and positions representing human intent based on joint torque sensors and robot dynamics models, reducing cost and installation complexity while enabling whole-body sensitivity
2. Torque-Based Contact Detection Mechanism: Introduce link-level localization to reduce optimization search space and enable real-time estimation
3. Contact Information-Based Adaptive Motion Planner: Infer human intent from contact information and replan robot motion online while maintaining smoothness and adapting to human corrections
4. Experimental Validation: Verify the accuracy of the force estimation method on a 7-DOF robotic arm and demonstrate the effectiveness of the adaptive motion planner under perceptual uncertainty

Methodology Details

Task Definition

Given an n-DOF serial manipulator, estimate the contact forces and positions representing human intent through joint torque measurements and robot dynamics models, and utilize this information for online motion planning adjustments.

Model Architecture

1. Dynamics Modeling

The robot dynamics equation incorporating external contact forces:

τ = M(q)q̈ + h(q,q̇) + τ_ext

where τ_ext represents joint torques induced by external contact forces.

2. Contact Detection

Detection statistic based on residual torques:

η(t) = ||W_τ τ̂_ext(t)||_2

Employ exponentially weighted moving average (EWMA) filtering and hysteresis mechanisms to ensure reliable state transitions.

3. Link-Level Localization

Determine contact link by analyzing residual torque distribution:

ℓ = max{j | |τ̂_ext,j| > τ_th and |τ̂_ext,j+1| < τ_th}

4. Contact Force Estimation

Model contact as point contact on a link, solved through constrained nonlinear optimization:

min_{s,F_c} (1/2)∑_{k=1}^N ||τ̂_ext(t_k) - J_c(q(t_k),s)^T F_c||^2
s.t. 0 ≤ s ≤ 1, ||F_c|| ≤ F_max_c

5. Adaptive Motion Planning

Adjust reference path based on estimated contact forces:

• Compute average contact force and direction
• Generate Cartesian deviation: Δx_c^(r) = α sat(||F̄_c^(r)||, F_max_c) f̂^(r)
• Employ C¹ polynomial bump functions to ensure path smoothness
• Accumulate deformations to preserve final target invariance

Technical Innovations

1. Whole-Body Sensitivity: Achieve contact sensing across the entire manipulator using existing joint torque sensors
2. Real-Time Performance: Reduce search space through link-level coarse localization, enabling real-time force estimation
3. Smooth Adaptation: Ensure C¹ continuity of path corrections using bump functions
4. Intent Inference: Directly map contact force magnitude and direction to motion correction intent

Experimental Setup

Experimental Platform

• Robot: 7-DOF Franka Research 3 manipulator
• Sensors: Joint torque sensors
• Environment: Computer disassembly collaborative scenario

Evaluation Metrics

• Force Estimation Accuracy: Assessed by comparing predicted joint torques with measured values
• Motion Adaptability: Evaluated through obstacle avoidance success rate and path smoothness
• Response Time: Latency from contact detection to motion adjustment

Experimental Design

1. Force Estimation Verification: Robot executes circular trajectory while human applies pushing, pulling, and tapping forces at different links
2. Single Contact Correction: Straight-line path obstacle avoidance scenario
3. Multiple Contact Correction: Complex obstacle avoidance requiring multiple human interventions

Experimental Results

Main Results

1. Force Estimation Accuracy

• Mean Absolute Error: 0.665 N·m (joint torque estimation)
• Temporal Consistency: Estimated torques align with measured values across all 7 joints
• Contact Detection: All 5 contact events detected timely

2. Motion Adaptation Performance

• Single Contact Correction: Successfully avoided vertically and horizontally placed obstacles
• Multiple Contact Correction: 4 consecutive contacts successfully guided robot around complex obstacles
• Path Smoothness: All corrected paths maintain C¹ continuity

Case Analysis

Experiment 1: Vertical Obstacle Avoidance

• Human applies horizontal negative x-direction contact force
• Robot successfully laterally offsets to avoid obstacle
• Path smoothly converges to original target point

Experiment 2: Horizontal Obstacle Avoidance

• Human applies upward contact force
• Robot navigates around obstacle from above
• Demonstrates correct understanding of intent in different directions

Experiment 3: Multiple Corrections

• 4 consecutive contacts: 2 lateral, 1 strong offset, 1 downward correction
• Demonstrates capability to handle complex intent sequences
• Validates effectiveness of incremental correction mechanism

Experimental Findings

1. Force Direction Consistency: Estimated force direction remains consistent within single contact events, indicating singular dominant intent
2. Force Magnitude Variation: Forces are smaller at contact onset and termination, peak in the middle, conforming to natural contact patterns
3. Real-Time Performance: System detects contact and adjusts motion in real-time, meeting collaboration requirements

Related Work

Major Research Directions

1. LLM Applications in HRC: High-level task planning and semantic reasoning
2. Physical Human-Robot Interaction: Kinesthetic guidance, impedance control, shared control
3. Contact Sensing Technologies: Tactile sensors, model-based force estimation
4. Intent Inference Methods: Machine learning, optimization methods, probabilistic models

Advantages of This Work

1. Cost-Effectiveness: Utilizes existing joint sensors without requiring additional tactile sensors
2. Real-Time Performance: Achieves real-time estimation through hierarchical localization strategy
3. Whole-Body Sensing: Enables full manipulator sensitivity compared to end-effector-limited approaches
4. Smooth Adaptation: Ensures path continuity, avoiding abrupt changes

Conclusions and Discussion

Main Conclusions

1. The proposed framework accurately estimates contact forces and positions representing human intent
2. Contact information-based motion planning achieves smooth online path adjustment
3. Experiments validate method effectiveness in collaborative disassembly tasks
4. Significantly reduces human operator burden compared to continuous kinesthetic guidance

Limitations

1. Point Contact Assumption: Current method assumes point contact, potentially unsuitable for large-area contact
2. Parameter Tuning: System sensitivity requires manual adjustment, may not adapt to individual differences
3. Environmental Constraints: Primarily validated in structured environments; complex environment adaptability insufficiently tested
4. Dynamics Model Dependency: Method relies on accurate robot dynamics models

Future Directions

1. Personalized Learning: Adapt to individual interaction styles through learning methods
2. Multiple Contact Handling: Extend to simultaneous multi-point contact scenarios
3. Environmental Generalization: Validate in more complex and unstructured environments
4. Long-Term Learning: Iteratively improve parameters through multiple trials

In-Depth Evaluation

Strengths

1. Technical Innovation: Cleverly combines dynamics modeling and optimization methods for low-cost, high-precision contact sensing
2. Practical Value: Addresses important problems in practical HRC with direct application potential
3. System Completeness: Forms complete closed-loop system from contact detection to motion planning
4. Comprehensive Experiments: Multi-scenario validation demonstrates robustness and adaptability

Weaknesses

1. Insufficient Theoretical Analysis: Lacks theoretical guarantees for convergence and stability
2. Limited Comparative Experiments: Insufficient quantitative comparison with other advanced methods
3. Missing User Studies: Lacks subjective evaluation from real users and long-term usage research
4. Computational Complexity: Insufficient analysis of computational overhead in optimization solving

Impact

1. Academic Contribution: Provides new technical pathway for intent inference in HRC
2. Industrial Application: Method shows promising prospects for industrial application
3. Reproducibility: Sufficient technical details facilitate reproduction and extension
4. Inspirational Value: Provides valuable insights for related research

Applicable Scenarios

1. Collaborative Manufacturing: Assembly, disassembly, polishing and other manufacturing tasks requiring human-robot collaboration
2. Service Robots: Scenarios requiring physical interaction in home and medical service applications
3. Rehabilitation Training: Rehabilitation robot applications requiring adaptation to patient intent
4. Educational Research: Research platform for HRC and intent inference

References

The paper cites 38 relevant references covering important works in HRC, pHRI, contact sensing, motion planning and other related domains, providing solid theoretical foundation for the research.

Overall Assessment: This is a high-quality paper with significant contributions to the human-robot collaboration field. The proposed adaptive motion planning framework based on contact information is technically advanced and highly practical, with comprehensive experimental validation. While there remains room for improvement in theoretical analysis and user studies, overall this represents important progress in the field with good academic value and application prospects.