whole-body control

To address the challenge of jointly ensuring motion stability and manipulation force control in legged mobile manipulation, this paper proposes a full-order inverse-dynamics-based whole-body model predictive control (MPC) framework. The method directly optimizes joint torques within a single prediction horizon, unifying whole-body motion planning and contact force generation to achieve dynamically consistent and constraint-complete natural coupling behavior. It integrates Pinocchio for rigid-body dynamics modeling, CasADi for automatic differentiation, and Fatrop for efficient interior-point optimization, enabling real-time control at 80 Hz on a Unitree B2 quadrupedal platform equipped with a Z1 manipulator. Experimental validation demonstrates robust performance across diverse dynamic manipulation tasks—including dragging heavy objects, pushing boxes, and wiping whiteboards—significantly enhancing both robustness and generalization capability of legged mobile manipulation.

Heavy-limb humanoid robots suffer from poor balance during dynamic locomotion and on unstructured terrain. To address this, we propose a model-based whole-body control framework integrating kinodynamics-aware reduced-order model predictive control (MPC) with a hierarchical quadratic programming (HQP) optimizer. For the first time, these two components are co-designed to explicitly capture time-varying center-of-mass and inertia distribution effects induced by heavy limbs, thereby enhancing whole-body coordination robustness. Contact force optimization and real-time motion planning enable dynamic walking at up to 1.2 m/s, rejection of external disturbances up to 60 N, and stable traversal of uneven ground and outdoor complex terrains. The method significantly improves balance maintenance and real-time adaptability of heavy-limb humanoids in dynamic and unstructured environments.

Coordinating whole-body locomotion and manipulation for quadrupedal robots equipped with robotic arms remains challenging due to tight coupling between legged locomotion and arm-end-effector dynamics. Method: This paper proposes RoboDuet, a dual-strategy synergistic framework enabling real-time locomotion-manipulation coupling. It introduces a novel decoupled–coupled architecture, integrating reinforcement learning training, full-state feedback modeling, hardware-in-the-loop optimization on real platforms, and cross-platform policy generalization—enabling zero-shot transfer across robots. Contribution/Results: RoboDuet overcomes limitations in 6-DoF end-effector pose tracking range and whole-body dynamic coordination inherent in conventional single-strategy approaches. In complex locomotion-manipulation (locomanip) tasks, it achieves a 23% improvement in success rate. The framework has been successfully deployed in real time and validated for generalization across multiple isomorphic quadrupedal robot platforms.

To address the challenge of achieving precise six-degree-of-freedom (6-DOF) end-effector (EE) pose tracking in task space for wheeled quadrupedal manipulator robots, this paper proposes a deep reinforcement learning (DRL)-based whole-body coordinated control framework. The method directly implements closed-loop 6D pose control in task space—bypassing hierarchical planning or inverse kinematics decomposition. Its key contributions are: (1) a nonlinear Reward Fusion Module (RFM) that explicitly models the multi-stage coupling among base motion, manipulator operation, and balance maintenance; and (2) a teacher–student hierarchical RL training paradigm to mitigate motion–balance coupling under high kinematic redundancy. Extensive simulation and real-robot experiments demonstrate smooth, robust tracking performance, achieving mean position error < 5 cm and orientation error < 0.1 rad—setting a new state-of-the-art.

Humanoid robots face significant challenges in maintaining dynamic balance and coordinating locomotion with manipulation when handling objects during walking, due to persistent or intermittent external contact forces—distinct from ground reaction forces. This paper proposes a dynamics-consistent locomotion-manipulation (locomanipulation) control framework. First, it embeds manipulation force reference trajectories directly into center-of-mass (CoM) motion planning; second, it designs a real-time feedback stabilizer based on external force tracking error. By integrating the linear inverted pendulum model with the divergent component of motion (DCM), the framework enables external-force-aware trajectory generation and robust gait adaptation. Evaluated in simulation and on a physical humanoid platform, the method significantly improves dynamic stability and trajectory tracking accuracy under complex manipulation tasks. It establishes a scalable theoretical and practical paradigm for unified locomotion-manipulation control.

This work addresses the challenges of high-dimensional action spaces, underutilized redundancy, and insufficient control accuracy in high-precision whole-body loco-manipulation for highly articulated robots. The authors propose a hierarchical control framework wherein a high-level planner leverages Kinematic Normalizing Flows to generate diverse, kinematically feasible partial reference trajectories in a latent space, effectively exploring redundant solutions. A low-level controller then employs imitation learning to accurately track these references while ensuring physical feasibility. By integrating a large-scale kinematic dataset with high-dimensional action modeling, the approach significantly outperforms existing methods in simulation. Hardware experiments across eight tasks and 24 trials demonstrate state-of-the-art performance, achieving end-effector pose errors of 4.5 cm and 0.14 rad, as well as mobile tracking errors of 0.1 m/s and 0.01 rad/s.

Existing whole-body control methods for large-scale humanoid robots often fail on long-tail motions involving high-dynamic transitions and critical balancing due to a mismatch between policy capabilities and motion demands. This work proposes Athena-WBC, a capability-aligned teacher–student framework that employs two specialized experts—one optimized for trajectory tracking under dynamic conditions and the other for training stability during balance-intensive tasks. The approach integrates routing-based distillation, DAgger imitation learning, constraint-aware objectives (without conservative cost shaping), gravity-based curricula, and reinforcement learning fine-tuning to distill these expert policies into a single deployable controller. Athena-WBC significantly enhances recovery robustness on long-tail motions and improves tracking performance on unseen actions, outperforming the strong baseline SONIC while requiring only a minimal set of expert policies.

This study addresses the challenge of enabling humanoid robots to perform full-body dance motions that simultaneously exhibit dynamism, stability, and artistic expressiveness. The authors propose a model-based two-stage framework: in the offline phase, motion capture data retargeting combined with trajectory optimization yields dynamically feasible dance sequences; in the online phase, a centroidal dynamics-based model predictive control (MPC) scheme employs long-horizon prediction to adjust footstep placements in real time for robust disturbance rejection. The approach is validated through the first public demonstration of robust, synchronized, four-minute dynamic dancing performed by multiple full-sized Kuavo 4Pro humanoid robots, showcasing its effectiveness in dynamic execution, disturbance resilience, and artistic fidelity.

This work addresses the challenge of enforcing runtime constraints—such as obstacle avoidance, joint limits, and center-of-mass stability—during deployment of reinforcement learning policies on humanoid robots. The authors propose ConstrainedMimic, a novel framework that, for the first time, integrates differentiable control barrier functions (CBFs) with full-body dynamics into a reinforcement learning-based tracking policy. By embedding kinematic and dynamic constraints directly into operational space control, the method enforces safety and feasibility in real time while minimally perturbing the original policy, and it flexibly accommodates new constraints post-training. The fully differentiable system supports deployment on CPU, GPU, or TPU, achieving 300–500 Hz real-time control in simulation on the Unitree G1 platform, and demonstrates high-fidelity whole-body motion tracking and teleoperation under self-collision and external obstacle avoidance, joint limit, and center-of-mass stability constraints.

This work addresses the limited whole-body coordination and generalization capability of legged robots when dynamically manipulating large, heavy objects in real-world environments. The authors propose a framework that integrates a pretrained whole-body control policy with a sampling-based online planner. By adapting the cost function at test time, the approach generalizes across diverse objects and tasks without requiring additional training, enabling flexible adjustment of behavioral objectives. The method is demonstrated on a Spot robot performing dynamic manipulation of oversized and overweight objects—such as uprighting heavy tires and dragging large barriers—for the first time. It is further validated in simulation on a humanoid robot executing tasks like opening doors and pushing tables, confirming its effectiveness and broad applicability.