Planning Human-Robot Co-manipulation with Human Motor Control Objectives and Multi-component Reaching Strategies

1. Problem Statement

Human-Robot Collaboration (HRC) requires robots to adapt to human intentions and actions in real-time. Existing approaches face significant limitations:

• Musculoskeletal Models: Effective for ergonomics (minimizing fatigue/torque) but limited to low-level functioning and lack generalizable behavior prediction.
• Data-Driven/Machine Learning Models: (e.g., Inverse Reinforcement Learning, Gaussian Processes) rely heavily on large datasets, which may be unavailable, impractical to collect, or lack generalizability to new scenarios.
• Gap: There is a lack of human motor control models that can provide generalizable behavior estimates, integrate into robot planning, and account for fundamental trade-offs like speed-accuracy and cost-benefit.

The core problem is to develop a robot planning framework that predicts human trajectories and intentions using principled computational models of human motor control, enabling seamless synchronization and authority handover without extensive training data.

2. Methodology

The authors propose a numerical trajectory optimization framework that embeds human motor control principles directly into the robot's planning cost function.

A. Human Motor Control Models

The framework integrates two key theoretical models:

1. Speed-Accuracy and Cost-Benefit Trade-offs:

   • Based on Fitts' Law and metabolic cost minimization.
   • The human trajectory minimizes an objective function $J$ that balances the benefit of reaching a goal (discounted by time) against the metabolic cost (joint torques).
   • Stochastic Dynamics: The model accounts for neural noise ($\epsilon$) that amplifies with muscle activity, leading to a distribution of possible trajectories rather than a single deterministic path.
   • Objective Function:
     $$J(\tau(t)) = E \int_0^\infty \left( e^{-t/\gamma} R(q, g) - \nu \|\tau(t)\|^2 \right) dt$$
     Where $R$ is the reward for reaching the goal, $\gamma$ is a time discount factor, and $\nu$ weights the metabolic cost.

2. Multi-Component Reaching Strategy:

   • Human reaching is modeled as two phases: an initial ballistic sub-movement (fast, less precise) and a final corrective sub-movement (slow, precise).
   • A transition point is predicted using a Gaussian Process Regression model trained on empirical data. This point depends on goal distance and width (difficulty).
   • This model is used to determine when a human would naturally switch from a "lead" role to a "fine-tuning" role.

B. Numerical Optimization Formulation

To make these continuous models tractable for robot planning:

• Discretization: The continuous dynamics are linearized and discretized using forward Euler integration.
• State Representation: The human state is treated as a Gaussian distribution $s \sim N(\mu, \Sigma)$ to handle noise.
• Reward Approximation: The binary goal indicator is replaced with a smooth, differentiable probability function (Gaussian kernel) allowing for closed-form expectation over the state distribution.
• Planning: The robot solves an optimal control problem to minimize the human-like cost function, generating trajectories that mimic human speed, acceleration, and noise characteristics.

C. Collaborative Applications

The framework is applied to two specific HRC scenarios:

1. Co-manipulation Synchronization: The robot predicts the human's trajectory in real-time (using observed partial motion) to synchronize its motion, matching velocity profiles and arrival times.
2. Authority Handover: The robot leads the initial ballistic phase of a task. Using the multi-component model, it predicts the transition point where the human would naturally begin corrective motion. At this point, the robot releases control (zero stiffness) to allow the human to take over for fine positioning.

3. Key Contributions

• Unified Motor Control Model: First integration of unified speed-accuracy and cost-benefit trade-offs into a robot planner for collaborative movements.
• Multi-Component Strategy Integration: Incorporation of ballistic/corrective movement splitting to determine optimal authority handover points.
• Goal Inference System: A method to estimate the human's remaining trajectory and goal location by constraining the optimization problem with observed partial motion.
• Validation: The approach is validated against Fitts' Law, velocity profile asymmetry, and dispersion characteristics, showing high fidelity to human motor control literature.

4. Results

The system was validated through computational simulations and physical experiments using a Franka Emika robotic arm in an electronics assembly task.

• Model Fidelity (O1):

  • The generated trajectories successfully reproduced the speed-accuracy trade-off: movements were faster for larger/shorter goals and slower for smaller/farther goals.
  • Velocity Profiles: The model produced temporally asymmetric velocity peaks (peak velocity occurring earlier in the movement), a characteristic observed in human motion but often missed by minimum-jerk models.
  • Cost-Benefit: The model allowed for a small probability of "missing" the goal to minimize time/energy, matching human risk-taking behavior.

• Co-manipulation Synchronization (O3):

  • In experiments where the robot and human moved an object together, the robot successfully synchronized its motion with the human.
  • The robot adapted to varying goal distances and directions in real-time.
  • Velocity profiles of the robot closely matched human profiles, reducing motion conflicts and interaction forces.

• Authority Handover (O2):

  • The robot successfully identified the transition point between ballistic and corrective phases.
  • Experimental data showed that the transition occurred near the point of maximum position dispersion (as predicted by the model).
  • The handover was smooth, with no sudden jumps in velocity, allowing the human to seamlessly take over for fine positioning.

5. Significance

• Principled vs. Data-Driven: This work demonstrates that robots can achieve human-like adaptability without massive datasets by leveraging established neuroscience principles (motor control theory).
• Natural Interaction: By mimicking human noise, speed, and strategy (ballistic vs. corrective), the robot's behavior becomes more "legible" and predictable to human partners, reducing cognitive load.
• Robustness: The approach handles uncertainty in goal location and varying task difficulties (distance/width) effectively, making it suitable for dynamic industrial environments like assembly lines.
• Future Impact: This framework provides a foundation for more intuitive HRC systems where robots can anticipate human needs and adapt their control authority dynamically, moving beyond simple pre-programmed paths or reactive impedance control.

Limitations & Future Work:
The current models are limited to point-to-point motions and do not fully account for complex interaction dynamics. The nonlinear solver currently takes >1 second (though linear versions are faster), and parameters require tuning. Future work includes user studies to assess subjective perceptions and extending models to continuous interaction dynamics.