Manip4Care: Robotic Limb Repositioning

• Manip4Care is an open-source, modular simulation pipeline for robotic limb manipulation, integrating physics simulation, antipodal grasp synthesis, and trajectory planning.
• It employs advanced MPPI control and vector-field execution to generate collision-free, dynamically optimal motions under strict biomechanical and environmental constraints.
• Evaluations demonstrate over 90% task success in assistive scenarios like bed bathing, highlighting its potential for personalized care, rehabilitation, and related applications.

Manip4Care is an open-source, modular simulation pipeline for robotic manipulation of human limbs, designed to enable robotic manipulators to grasp and reposition limbs effectively during assistive care tasks. Its core objective is to address the limitations of previous assistive robotics solutions that assume the human target remains static or quasi-static, thereby expanding the domain of robotic assistance to include dynamic, constrained limb repositioning as required in tasks such as bed bathing and dressing. The Manip4Care pipeline integrates advanced grasp synthesis, trajectory generation, and strict enforcement of biomechanical and collision avoidance constraints, targeting applications in personal care, rehabilitation, and related scenarios (Koh et al., 4 Aug 2025).

1. System Architecture and Pipeline Structure

Manip4Care is architected as a modular pipeline progressing from task specification to execution, with distinct modules for physics simulation, grasp generation, and constrained trajectory planning. The pipeline is typically instantiated with a downstream task (e.g., repositioning a limb to expose regions for bed bathing) and proceeds via:

• Physics Simulation Environment: Built using PyBullet, providing rigid-body dynamics, continuous collision detection, and integrated forward/inverse kinematics for both the robot and a realistic human limb model.
• Grasp Generation: Employs antipodal sampling on limb surface point clouds with surface normal computation and force closure validation to synthesize robust, feasible grasps.
• Trajectory Planning: Leverages a Model Predictive Path Integral (MPPI) control framework, iteratively generating, perturbing, and scoring collision-free motion plans subject to strict biomechanical and environmental constraints.
• Vector-Field-Based Execution: Follows the optimal trajectory using a strategy that combines attraction along the planned trajectory with repulsion generated from real-time collision checks.
• Closed-loop Correction: An iterative "IK-handshaking" algorithm ensures that the robot–human kinematic loop constraints are satisfied at each step of trajectory execution.

This architecture is designed to flexibly integrate with downstream assistive tasks (for example, pairing with a wiping robot in bed bathing scenarios), and to generalize across different initial conditions and limb postures.

2. Grasp Synthesis Using Antipodal Sampling and Force Closure

A central capability of Manip4Care is its grasping method, which employs antipodal sampling combined with force closure constraints to robustly grasp human limbs. The process begins with:

• Point Cloud Acquisition and Normal Estimation: A dense point cloud of the limb is generated, and surface normals are computed for each point.
• Antipodal Pair Selection: Candidate grasp pairs are identified as points with surface normals approximately opposed in direction.
• Force Closure Verification: For each candidate, friction-cone overlaps at the contact points are verified to ensure no-slip, balanced forces.
• Grasp Scoring and Selection: Candidate grasps are scored using the metric

$S_i = \alpha d_i - (1-\alpha) r_i$

where $d_i$ is the Euclidean distance between the candidate location $p_{r_i}$ and a reference location $p^H_{ref}$, $r_i$ is the inner product of rotation quaternions measuring orientation agreement, and $\alpha \in [0,1]$ weights the two terms. Only grasps that are kinematically feasible, collision-free, and sufficiently aligned with the limb's principal axis (and at a safe distance from joints) are retained.

The grasping strategy thus ensures both physical robustness (force closure) and practical suitability (avoidance of sensitive anatomical regions).

3. Trajectory Planning with MPPI and Constraint Enforcement

Manip4Care’s trajectory planner operates after a rigid grasp is established between the robot’s end effector and the limb:

• Nominal Control Sequence Generation: The planner interpolates between the current and goal limb configurations.
• MPPI Stochastic Planning: Nominal sequences are perturbed stochastically (Gaussian noise, $\lambda=1$, small covariance $\Sigma$), and each perturbed plan is refined with an iterative IK-handshaking method to satisfy the kinematic loop constraint between robot and limb.
• Constraint Checking: Each candidate trajectory is scored based on goal convergence, trajectory length, and collision penalties. Continuous collision checking is performed using a Configuration Signed Distance Function (CSDF) computed over sets of control points on robot and limb, enforcing a safety margin $\rho=2$ cm and outer penalty radius $r=5$ cm.
• Closed-Loop Kinematics: The robot and the human limb are each treated as kinematic chains with shared closed-loop constraints at the fixed grasp. For each planned motion, forward kinematics and inverse kinematics are recomputed for both chains to maintain the rigid grasp constraint.

At execution time, a local vector-field controller jointly drives the end effector along the trajectory while dynamically avoiding collisions by synthesizing corrective repulsion forces from the CSDF gradients. At each step, IK-handshaking projects the motion onto the joint configuration manifold.

4. Biomechanical and Environmental Constraint Integration

Critical to safe human–robot interaction, Manip4Care integrates both anatomical and environmental constraints into planning and execution:

• Biomechanical Joint Limits: Human limb motion is modeled with clinically-sourced feasible joint ranges for individual degrees of freedom (shoulder flexion, external rotation, abduction, elbow angle). At all stages in planning and execution, the system verifies that the human limb configuration $q^H_t \in Q^H$ for $t \in [0, T]$ does not violate joint or anatomical constraints.
• Collision Avoidance: All candidate plans and online trajectories are subject to real-time collision detection using the CSDF for both human and robot bodies. Control points sampled on both the robot and the limb are checked for proximity violations against a discretized representation of the environment and other agents.
• Closed-Loop Constraint Enforcement: The grasp is modeled via fixed transformation matrices $T^{cp}_{eef}$ (robot end effector in grasp frame) and $T^{cp}_{limb}$ (human limb in grasp frame). These are enforced throughout the planning and correction process to maintain closed-chain kinematic consistency.

The mathematical model governing configuration feasibility is:

• All $q_t^R \in Q^R$, $q_t^H \in Q^H$ at each time $t$ (configuration spaces for robot and human, respectively),
• Satisfying closed-loop constraint dictated by the grasp transform.

5. Evaluation: Scenarios, Metrics, and Results

Manip4Care was evaluated on canonical assistive care tasks:

• Task Scenarios: Repositioning an upper limb in both supine and sitting positions, using a 4-DOF upper limb model. Evaluations include both simulated (UR5 robot) and real-world (mannequin) implementations.
• Biomechanical Range Comparisons: Plan and execution feasibility were analyzed across different age brackets with varying joint limits (e.g., reduced shoulder mobility in older subjects).
• Success Metrics:
  • Success rate: Over 90% in supine/sitting tasks.
  • Joint-limit violations: Near zero, indicating reliable adherence to anatomical safety.
  • Planning time: ~3–4 s per trajectory.
  • Total execution time: Generally under 10 s.
• Case Study—Bed Bathing: Integration with a downstream wiping robot task demonstrates that combining active limb repositioning with wiping leads to substantially increased coverage of target regions versus static human poses. In this context, next-arm configurations chosen by a neural net policy yielded higher task success and lower execution time than random sampling, despite the latter sometimes producing marginally higher coverage.

A plausible implication is that such modular integration of manipulation and trajectory planning with biomechanical constraint enforcement constitutes a robust framework adaptable to a broad range of assistive robotics tasks.

6. Limitations, Future Directions, and Broader Application

The current development assumes a passive human limb model (no active resistance/assistance by the human); future work aims to incorporate active limb dynamics, such as implicit muscle torques, for more realistic simulation of human–robot interaction. The limb is presently modeled as a rigid kinematic chain; research extensions may introduce soft-body dynamics for more accurate modeling of limb compliance and skin deformation.

Broader applications for Manip4Care include robotic dressing, physiotherapeutic exercise guidance, and complex search-and-rescue scenarios where human-safe manipulation is required. The modular pipeline is positioned as a reference implementation facilitating further research and development in physically sensitive, constraint-driven assistive robotics for human care.

7. Summary and Significance

Manip4Care represents an advance in robotic assistive care by enabling safe, autonomous, and anatomically-constrained limb repositioning in human–robot shared environments. The integration of antipodal, force-closure-driven grasp synthesis, stochastic MPPI trajectory optimization, vector-field-based collision avoidance, and strict biomechanical constraint adherence distinguishes the framework from static-pose approaches. Experimental demonstrations, including simulated and real-world settings with varying anatomical constraints and downstream task integration, validate its effectiveness and robustness. As future extensions incorporate a broader range of human dynamics and environmental factors, Manip4Care is poised to support increasingly complex and personalized assistive care tasks (Koh et al., 4 Aug 2025).