[Paper Review] Goal Set Inverse Optimal Control and Iterative Re-planning for Predicting Human Reaching Motions in Shared Workspaces

Jim Mainprice, Rafi Hayne+1 more|arXiv (Cornell University)|Jun 7, 2016

Robot Manipulation and Learning33 references4 citations

TL;DR

This paper proposes a Goal Set Inverse Optimal Control (IOC) framework combined with iterative re-planning to predict human reaching motions in shared workspaces. By learning a cost function from human motion capture data and using it in a STOMP-based planner that re-plans dynamically over a goal region, the method outperforms hand-tuned baselines in predicting both human-human and human-robot collaborative reaching tasks, with improved accuracy as humans acclimate to the robot's motion.

ABSTRACT

To enable safe and efficient human-robot collaboration in shared workspaces it is important for the robot to predict how a human will move when performing a task. While predicting human motion for tasks not known a priori is very challenging, we argue that single-arm reaching motions for known tasks in collaborative settings (which are especially relevant for manufacturing) are indeed predictable. Two hypotheses underlie our approach for predicting such motions: First, that the trajectory the human performs is optimal with respect to an unknown cost function, and second, that human adaptation to their partner's motion can be captured well through iterative re-planning with the above cost function. The key to our approach is thus to learn a cost function which "explains" the motion of the human. To do this, we gather example trajectories from pairs of participants performing a collaborative assembly task using motion capture. We then use Inverse Optimal Control to learn a cost function from these trajectories. Finally, we predict reaching motions from the human's current configuration to a task-space goal region by iteratively re-planning a trajectory using the learned cost function. Our planning algorithm is based on the trajectory optimizer STOMP, it plans for a 23 DoF human kinematic model and accounts for the presence of a moving collaborator and obstacles in the environment. Our results suggest that in most cases, our method outperforms baseline methods when predicting motions. We also show that our method outperforms baselines for predicting human motion when a human and a robot share the workspace.

Motivation & Objective

• To enable accurate prediction of human reaching motions in collaborative, shared workspaces for safer and more efficient human-robot interaction.
• To model human motion as optimal with respect to an unknown cost function, learned from observed trajectories.
• To capture human adaptation to a collaborator’s motion through iterative re-planning using the learned cost function.
• To evaluate the method in both human-human and human-robot collaborative settings, particularly under dynamic interference.
• To demonstrate that inverse optimal control with goal set re-planning yields better motion predictions than hand-tuned cost functions.

Proposed method

• Collect motion capture data from human pairs performing a collaborative assembly task, segmenting into individual reaching motions.
• Use Path Integral Inverse Reinforcement Learning (PIIRL) to learn a cost function as a weighted sum of feature functions encoding smoothness and inter-agent distances.
• Adapt the STOMP trajectory optimizer for iterative re-planning in dynamic environments with moving collaborators and obstacles.
• Modify the planner to search over a task-space goal region rather than a fixed goal point, improving robustness to uncertainty.
• Map the same feature functions to robot kinematics to enable prediction in human-robot shared workspace scenarios.
• Evaluate predictions using Dynamic Time Warping (DTW) between predicted and demonstrated trajectories in both joint and task space.

Experimental results

Research questions

• RQ1 Can inverse optimal control with iterative re-planning accurately predict human reaching motions in collaborative shared workspaces?
• RQ2 Does learning a cost function from human demonstrations outperform hand-tuned cost functions in predicting human motion?
• RQ3 How does the prediction accuracy of the proposed method change as humans acclimate to a robot’s motion in a shared workspace?
• RQ4 To what extent does iterative re-planning better capture human adaptation to a collaborator’s motion compared to single-shot planning?
• RQ5 Can the same cost function learned from human-human interactions be effectively applied to predict human motion in human-robot collaboration?

Key findings

• The IOC-based method outperformed both baseline tuning methods (baseline 0 and baseline 1) in predicting human motion, as measured by lower DTW scores in both task-space and joint center distances.
• The distribution of DTW scores for the IOC method showed consistent improvement over baselines across all 2120 human reaching motions in the human-robot experiment.
• Prediction accuracy improved significantly over time: the mean and standard deviation of DTW scores for the IOC method decreased from the first to the seventh run, indicating better adaptation as humans acclimated to the robot.
• Outliers in the IOC distribution were primarily due to human hesitations or collision-free pathfinding failures within iteration limits, suggesting limitations in handling non-continuous motion.
• Iterative re-planning was more effective than single-shot planning in capturing human adaptation when motion interference was significant.
• The method successfully generalized the cost function learned from human-human data to predict human motion in human-robot collaboration, demonstrating transferability.