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

For successful goal-directed human-robot interaction, the robot should adapt to the intentions and actions of the collaborating human. This can be supported by musculoskeletal or data-driven human models, where the former are limited to lower-level functioning such as ergonomics, and the latter have limited generalizability or data efficiency. What is missing, is the inclusion of human motor control models that can provide generalizable human behavior estimates and integrate into robot planning methods. We use well-studied models from human motor control based on the speed-accuracy and cost-benefit trade-offs to plan collaborative robot motions. In these models, the human trajectory minimizes an objective function, a formulation we adapt to numerical trajectory optimization. This can then be extended with constraints and new variables to realize collaborative motion planning and goal estimation. We deploy this model, as well as a multi-component movement strategy, in physical collaboration with uncertain goal-reaching and synchronized motion tasks, showing the ability of the approach to produce human-like trajectories over a range of conditions.

💡 Research Summary

The paper presents a novel framework that integrates well‑established human motor control principles directly into robot trajectory planning for human‑robot co‑manipulation. Existing approaches either rely on musculoskeletal models that address only low‑level ergonomic concerns or on data‑driven machine‑learning methods that require large amounts of interaction data and often lack generalizability. In contrast, the authors adopt two fundamental human motor control trade‑offs – the speed‑accuracy relationship (captured by Fitts’ law) and the cost‑benefit balance between metabolic effort and task benefit – and formulate them as a continuous optimal‑control objective.

The core objective function is
(J(\tau(t)) = \int_{0}^{\infty} e^{-t/\gamma} R(q,g) - \nu |\tau(t)|^{2} , dt),
where (R(q,g)) is a reward that reflects whether the end‑effector lies within a goal region of radius (W). To make the reward differentiable and compatible with stochastic dynamics, the authors replace the binary indicator with a Gaussian probability density centered on the goal, yielding a smooth reward that depends on the current mean state and covariance.

Human arm dynamics are modeled by the standard torque equation (M(q)\ddot q + D(\dot q) + G(q) = \tau (I + \epsilon)), where (\epsilon) is zero‑mean Gaussian neural noise whose variance scales with muscle activation. The continuous dynamics are discretized using forward Euler and linearized, resulting in a state‑space update (s_{k+1}=A s_k + B \tau_k). The state (s_k) (joint positions and velocities) is treated as a Gaussian random variable with mean (\mu_k) and covariance (\Sigma_k), which are propagated analytically through the linear dynamics.

A second contribution is the explicit modeling of the well‑known two‑phase reaching strategy: an initial ballistic sub‑movement that is fast but imprecise, followed by a corrective sub‑movement that is slower and more accurate. The transition point between these phases depends on goal distance and width; the authors learn this dependency from existing human reaching data using Gaussian‑process regression. In collaborative tasks, the robot can use the predicted transition point to decide when to hand over authority to the human, i.e., when the human is expected to switch from ballistic to corrective control.

The planning problem becomes a finite‑horizon optimal‑control problem:
(\min_{\tau_{0:H}} \sum_{i=0}^{H} \gamma^{-i} R(\mu_i,\Sigma_i) + \nu |\tau_i|^2)
subject to the discretized dynamics and initial conditions. Because the reward is a closed‑form expectation over Gaussian states, the objective is smooth and can be solved with standard nonlinear programming solvers.

Experimental validation is performed in two scenarios. First, a planar point‑to‑point reaching task with varying goal widths (0.02–0.05 m) and distances (0.2–0.8 arm lengths) demonstrates that the generated velocity profiles and movement times match Fitts’ law predictions and reproduce the dispersion patterns observed in human experiments. Second, a collaborative manipulation task requires the robot to execute the high‑speed portion of the motion while the human takes over for the fine‑positioning phase. The robot uses the learned transition model to trigger an impedance‑based handover; human participants show reduced positioning error during the corrective phase, confirming that the handover occurs at a natural point in the movement.

The authors argue that their approach offers three main advantages: (1) it leverages fundamental, well‑validated human motor control theory, providing behavior predictions that generalize across tasks without extensive training data; (2) it naturally accommodates additional constraints such as goal uncertainty, role switching, or safety limits; and (3) it yields human‑like trajectories that improve legibility and predictability for the human partner. Limitations include the reliance on linear‑Gaussian approximations, which may not capture highly nonlinear musculoskeletal dynamics or complex multi‑joint interactions. Future work is suggested to extend the model to nonlinear dynamics, incorporate real‑time sensory feedback for online parameter adaptation, and explore multi‑human collaboration scenarios.

Overall, the paper demonstrates that embedding human motor control objectives into robot planning bridges the gap between purely ergonomic models and data‑heavy learning approaches, enabling more intuitive and efficient human‑robot co‑manipulation.