Whole-Body Manipulation

• Whole-Body Manipulation (WBM) is the coordinated control of whole-body functions like robot locomotion, manipulator, joint, contact-force, torque and environment interactions as fifth-system.
• The primary setup of the whole-body platform involves mobile manipulators, arrayed manipulators, humanoids, soft robots, bipes, and quadrupedal while the method is applied through Nonlinear Model-Predictive Control (MPC) and admittance.
• Some variants of Whole-Body Manipulation include arm motion, postural adaptation, stepping, and required support according to task space

Whole-body manipulation (WBM) is the coordinated control or learning of a robot’s manipulator, locomotion system, torso, support contacts, and environmental interactions as a coupled system. Unlike end-effector-only manipulation, WBM treats base motion, body posture, contact formation, balance, actuator effort, and object interaction as potentially useful manipulation variables. Depending on the platform and methodology, WBM is formulated through nonlinear model-predictive control, whole-body inverse dynamics, hierarchical optimization, reinforcement learning, teleoperation, motion retargeting, or combinations of these approaches. The common objective is to exploit the robot’s complete embodiment to extend reachability, regulate interaction forces, manipulate large or ungraspable objects, and maintain dynamic feasibility.

1. Conceptual foundations and system-level formulation

WBM differs from sequential architectures in which an end-effector trajectory is first planned, converted into joint commands, and subsequently stabilized by an independent locomotion or balance controller. In a whole-body formulation, the arm and mobile base, or the arms, torso, legs, and feet, are decision variables of a common control or policy layer. The robot may reduce an end-effector error by moving the arm, translating or rotating the base, tilting the body, changing support contacts, or combining these motions.

For an underactuated or dynamically unstable platform, the coupling is essential. Arm acceleration changes the whole-body center of mass, angular momentum, inertia, and reaction forces. Environmental contact introduces additional forces and moments that can destabilize the base. A unified controller therefore evaluates manipulation actions according to their predicted influence on future balance, actuator requirements, friction feasibility, and support configuration.

The formulation varies substantially across robot classes:

• Mobile manipulators combine an arm with a wheeled or balancing base. The base can enlarge the manipulator workspace, compensate for arm-induced disturbances, or contribute directly to end-effector motion.
• Legged manipulators use locomotion, body posture, foothold selection, and arm motion jointly. Their whole-body degrees of freedom can support manipulation on uneven terrain.
• Humanoids can contact objects with hands, forearms, elbows, torso, chest, or legs, allowing non-prehensile and multi-contact manipulation.
• Quadrupeds can use the body itself to push and reorient objects that cannot be grasped.
• Soft robots exploit compliance and distributed contact to envelop bulky objects.
• Bipedal mobile manipulators can infer whether an end-effector target requires arm motion, postural adaptation, or stepping without explicit base or footstep commands.

The underlying task can be represented as a coupled mapping from task-space objectives to whole-body actions:

$$\text{task target} \rightarrow \text{body configuration and contact strategy} \rightarrow \text{whole-body actuation}.$$

The task target may be an end-effector pose, an interaction wrench, an object pose, a locomotion command, or a natural-language instruction. WBM does not require every method to model the manipulated object explicitly. Some approaches optimize object and contact dynamics directly, while others specify only robot-side forces and velocities or infer object behavior from interaction history.

2. Model-based control and predictive optimization

Whole-body MPC for unstable mobile manipulators

A representative model-based formulation is the nonlinear whole-body MPC developed for the ETH Zurich Rezero ballbot, a roughly $25\,\mathrm{kg}$ robot balancing on a single spherical wheel and equipped with a torque-controlled three-degree-of-freedom arm (Minniti et al., 2019). The system has eight generalized degrees of freedom and six independent actuation inputs, making it underactuated. Its model includes the ball, body orientation, arm joints, actuator torques, and end-effector interaction wrench.

The controller optimizes manipulation, balance, and environmental interaction in one finite-horizon problem. Its state is

$x=\begin{bmatrix}q\v\end{bmatrix},$

and its input contains actuator torques and end-effector wrench:

$$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$

The full nonlinear dynamics include the whole-body mass matrix, nonlinear effects, actuator-selection mapping, end-effector Jacobian, and the generalized force induced by the contact wrench:

$\dot{x}=f(x,u).$

The optimization is transcribed at the end-effector level: references are specified as end-effector position, orientation, velocity, and wrench objectives, while the optimizer retains the complete generalized state, nonlinear dynamics, and actuator torques internally. Consequently, an end-effector error can be reduced by arm motion, ball translation, body rotation, body tilt, or a combination of these. The optimizer can also accept end-effector error when exact tracking would compromise future balance or actuator feasibility.

The end-effector wrench is treated as an optimization input. A desired wrench is imposed through a weighted tracking term, and a friction cone constrains the contact force. In door-pushing experiments, desired normal forces of $5\,\mathrm{N}$ and $10\,\mathrm{N}$ were considered with friction coefficient $\mu=0.7$. The larger force produced a larger body inclination, illustrating that the robot uses its whole body to generate interaction force.

The method uses Sequential Linear Quadratic MPC with real-time iteration. Each update forward-integrates the nonlinear dynamics, linearizes the trajectory, quadratizes the cost, solves a constrained time-varying LQR problem, performs a line search, and warm-starts the next horizon. The experiments used an MPC frequency of approximately $70\,\mathrm{Hz}$, state estimation at $400\,\mathrm{Hz}$, low-level motor tracking at $25\,\mathrm{kg}$0, and horizons between $25\,\mathrm{kg}$1 and $25\,\mathrm{kg}$2. The controller was validated in hardware for end-effector pose tracking, disturbance rejection, coordinated ball and arm motion, and door opening.

Inverse-dynamics MPC for legged loco-manipulation

Whole-body inverse-dynamics MPC extends this principle to legged robots by optimizing configuration, velocity, acceleration, contact forces, and joint torques simultaneously (Molnar et al., 24 Nov 2025). The floating-base dynamics are enforced through a contact-aware recursive Newton–Euler equation:

$25\,\mathrm{kg}$3

The contact-force vector includes forces at the feet and the arm end effector. For the Unitree B2 equipped with a Z1 arm, the reported model has $25\,\mathrm{kg}$4 degrees of freedom: six floating-base degrees of freedom, twelve leg joints, and four active arm joints.

The arm’s end-effector force and velocity can be imposed as explicit constraints rather than merely soft penalties. This permits simultaneous force and velocity tracking while allowing the base, legs, contact forces, and foot trajectories to adapt. Stance feet satisfy zero-velocity and friction constraints; swing feet have zero contact force and a prescribed vertical velocity profile. Horizontal swing-foot motion remains available to the optimizer, allowing manipulation demands to influence foot placement.

The implementation uses Pinocchio, CasADi, and Fatrop. The compiled controller runs at $25\,\mathrm{kg}$5, with approximately $25\,\mathrm{kg}$6 average solve time in the reported benchmark. Hardware demonstrations include pulling a $25\,\mathrm{kg}$7 load, pushing a box against a wall while trotting in place, and wiping a whiteboard. The controller provides robot-side force and motion coordination, but the manipulated object is not modeled and the experiments do not report quantitative force-tracking measurements.

Contact optimization

Contact-rich WBM introduces both continuous and discrete complexity. Contact may be created, maintained, slid, or broken; the contact location may lie anywhere on the robot or object surface; and different body links may be selected. “Explicit Contact Optimization in Whole-Body Contact-Rich Manipulation” addresses planar manipulation by representing robot and object outlines with continuous differentiable surface parameters (Leve et al., 2024). Contact locations are optimized through continuous coordinates rather than selected from a dense finite sample.

The planner combines randomized selection of tree nodes and contacting links with continuous optimization of contact locations, robot configurations, contact impulses, sliding velocities, and trajectories. It uses contact-maintenance equality, Coulomb friction, complementarity constraints, collision avoidance, quasi-dynamic transitions, and long-horizon guide tracking. The hierarchy consists of context sampling, short-horizon contact planning, long-horizon guide generation, and guide tracking.

On planar reorientation benchmarks, the method achieved approximately $25\,\mathrm{kg}$8 fewer iterations and $25\,\mathrm{kg}$9 less planning time than the considered sampling-based comparison on average, while maintaining a $x=\begin{bmatrix}q\v\end{bmatrix},$0 success rate across the tested angles. The formulation remains non-convex, planar, quasi-dynamic, restricted to one point of contact at a time, and too slow for real-time MPC; hardware replanning required approximately three minutes.

A related approach uses rounded segment patches with closed-form proximity computations for robot and object surfaces (Levé et al., 18 Aug 2025). Its cost combines contact proximity, object-directed force quality, and robot-side manipulability. The planner can discover single-arm, dual-arm, torso, contact-switching, and low-degree-of-freedom body-contact strategies. In planar benchmarks, it achieved a $x=\begin{bmatrix}q\v\end{bmatrix},$1 average planning-time improvement over the state-of-the-art sampling planner and was transferred to humanoid hardware for manipulation involving fingers, wrists, elbows, torso, and one or both arms.

3. Whole-body kinematics, redundancy, and interaction control

Augmented kinematics

For a mobile manipulator, whole-body kinematics are commonly expressed through an augmented Jacobian. In passive collaborative transportation, the end-effector spatial velocity is

$x=\begin{bmatrix}q\v\end{bmatrix},$2

with

$x=\begin{bmatrix}q\v\end{bmatrix},$3

The base contribution incorporates the nonholonomic differential-drive structure and an input-output feedback-linearized control point. A Moore–Penrose pseudoinverse maps the desired operational velocity to wheel and arm velocities:

$x=\begin{bmatrix}q\v\end{bmatrix},$4

The base and arm are therefore coordinated through one whole-body Jacobian rather than independent inverse-kinematics controllers (Benzi et al., 2022). Control barrier functions can suppress arm motion near a workspace boundary, allowing the base to contribute more strongly.

Admittance, passivity, and collaborative interaction

The same system uses measured human interaction forces as input to an adaptive operational-space admittance. The desired end-effector motion is generated from time-varying inertia and damping, with damping reduced during acceleration and increased during deceleration. Since online parameter variation can inject energy into the human–robot interaction, an energy tank passivizes the admittance.

The tank has storage function

$x=\begin{bmatrix}q\v\end{bmatrix},$5

and must remain above a nonzero lower bound. A passive-approximation optimization modifies the desired admittance velocity only as much as necessary to preserve the tank energy constraint. Control barrier functions impose obstacle-clearance, self-collision, and manipulator-workspace constraints, while slack variables preserve numerical feasibility when constraints conflict.

The experimental platform was a MiR 100 differential-drive base with a UR10e six-degree-of-freedom arm transporting a $x=\begin{bmatrix}q\v\end{bmatrix},$6 wooden plank with a human. The adaptive controller produced visibly lower human effort than fixed admittance, maintained the $x=\begin{bmatrix}q\v\end{bmatrix},$7 capsule safety distance, and allowed the base to continue moving when arm motion was halted at workspace boundaries. The guarantees apply primarily to the commanded kinematic and energy-balance models; they do not establish passivity of every unmodeled hardware, sensing, delay, saturation, and human–load interaction effect.

Teleoperation and dynamic force regulation

Whole-body bilateral teleoperation maps human body posture, center of pressure, arm motion, and environmental interaction to a wheeled humanoid (Purushottam et al., 2023). The robot SATYRR is represented as a wheeled cart-pole, while the operator is represented by a fixed-base actuated inverted pendulum. Human pitch, pitch rate, and center of pressure are mapped to robot dynamic locomotion. Haptic feedback communicates dynamic mismatch, robot–environment force, and a virtual leaning spring.

The system enables the operator to simultaneously regulate robot pitch, wheel force, contact location, and arm motion. In heavy-box pushing, the robot moved boxes at approximately $x=\begin{bmatrix}q\v\end{bmatrix},$8, including boxes up to approximately $x=\begin{bmatrix}q\v\end{bmatrix},$9 of the robot’s stated reduced-order center-of-mass mass. A moving-object handoff involved grasping a $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$0 package while moving at approximately $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$1. The experiments also show limitations: force estimates were noisy, the operator had difficulty with fine force discrimination, and the pushing experiments were planar and constrained by rails.

Whole-body force and object modeling

Model-based approaches differ in whether they model the manipulated object. The inverse-dynamics MPC formulation specifies robot-side end-effector force and velocity but does not predict object motion. The contact-optimization methods explicitly model object pose, contact points, friction, and quasi-dynamics. Learned systems may omit object models and infer object properties from temporal interaction history.

Quadrupedal whole-body pushing provides an example of implicit object adaptation (Jeon et al., 2023). A hierarchical policy controls planar body velocity, while a recurrent latent encoder infers manipulation-relevant properties from interaction history, proprioception, and action history. The deployment policy does not require explicit mass, inertia, friction, center-of-mass, or drag estimates. In simulation, the method achieved $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$2 success under $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$3 and $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$4 tolerances; real-world experiments achieved $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$5 successful trials on objects including a $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$6 water-filled drum manipulated by a $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$7 quadruped.

4. Learning-based whole-body coordination

Hierarchical and cooperative policies

RoboDuet separates locomotion and manipulation into two interacting policies (Pan et al., 2024). The locomotion policy controls twelve leg joints, while the arm policy controls six arm joints and outputs desired quadruped pitch and roll. The arm therefore requests body-posture changes, and the locomotion policy realizes them while maintaining balance. Training occurs in two stages: locomotion is first trained with the arm fixed in a default configuration, then the arm policy is trained on top of the locomotion policy.

The system achieves large-range six-degree-of-freedom end-effector tracking and zero-shot reuse of the arm policy across quadrupeds with similar morphology, including Unitree Go1, A1, and Go2 platforms. The reported moving workspace for Go1 plus Arx5 was approximately $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$8, compared with $$u=\begin{bmatrix}\tau\\lambda_{ee}\end{bmatrix}.$$9 for the baseline. Transfer still requires training a locomotion policy for the new quadruped.

A bipedal mobile-manipulation controller instead maps a single six-degree-of-freedom end-effector target to all fourteen actuated joints of a LimX TRON 1 biped with an ARX L5 arm (Chen et al., 16 Sep 2026). No base velocity, gait phase, footstep, or discrete reach-versus-step command is provided. The policy learns whether to move the arm, alter posture, squat, or step. A phase-aware reward gate shifts emphasis between approach and precision manipulation, while a Transformer–GRU temporal context estimator predicts base velocity and latent dynamics information from observation history.

Terrain-aware legged manipulation

Terrain-Aware Whole-Body Control integrates foot-centric exteroception, online physical-parameter adaptation, arm inverse kinematics, and RL-based leg control (Guo et al., 29 May 2026). The policy receives ten-step proprioceptive history and a $\dot{x}=f(x,u).$0 foot-centric elevation representation. Ring-wise convolutions encode terrain around each foot. The arm uses damped-least-squares inverse kinematics, while the learned leg policy changes body height, pitch, roll, and support geometry to make arm targets reachable.

End-effector targets are sampled relative to a plane fitted through the four foot contacts. This foot-contact-plane representation decouples manipulation targets from base fluctuations and terrain inclination. A dual-policy distillation procedure preserves manipulation behavior from a quasi-level-terrain policy while adding terrain-adaptive locomotion.

On complex terrain, TA-WBC achieved a reachable workspace of $\dot{x}=f(x,u).$1, end-effector error of $\dot{x}=f(x,u).$2, unexpected-contact rate of $\dot{x}=f(x,u).$3, and maximum tested terrain level $\dot{x}=f(x,u).$4. The real robot performed a long-horizon sequence involving bottle pickup, climbing five $\dot{x}=f(x,u).$5 stairs, throwing the bottle into a bin, descending, and returning to the starting region. The task succeeded five consecutive times without falls or stumbles.

Human motion, teleoperation, and demonstration interfaces

Learning systems increasingly use human motion to acquire whole-body data. Astribot Suite combines a $\dot{x}=f(x,u).$6-degree-of-freedom mobile dual-arm robot, a Meta Quest 3S interface, whole-body control, and a transformer-based conditional diffusion policy (Gao et al., 23 Jul 2025). The learned policy uses RGB observations and egocentric end-effector delta actions, while a real-time trajectory-generation module blends asynchronous action chunks to enforce smoothness and velocity limits.

Across six household tasks, the reported average success rate was approximately $\dot{x}=f(x,u).$7, with some tasks reaching $\dot{x}=f(x,u).$8. The most difficult component was pressing a small trash-bin button, which achieved $\dot{x}=f(x,u).$9 success. The method demonstrates the importance of embodiment, interface design, end-effector-space action representations, and post-processing between neural predictions and hardware execution.

HumanoidExo transfers human upper-body motion through a wearable exoskeleton, while lower-body behavior is represented by forward velocity, yaw rate, and torso-height commands (Zhong et al., 3 Oct 2025). Direct joint-space transfer is used for much of the arm, iterative inverse kinematics is used for shoulder and wrist alignment, and LiDAR estimates torso motion. Five direct robot demonstrations combined with $5\,\mathrm{N}$0 exoskeleton demonstrations achieved approximately $5\,\mathrm{N}$1 on tabletop manipulation and $5\,\mathrm{N}$2 on seen-cloth whole-body manipulation. Walking was acquired solely from exoskeleton data in the reported walk-and-place experiment.

HuMI extends robot-free data collection using handheld grippers, wrist cameras, pelvis and foot trackers, human-in-the-loop inverse-kinematics adaptation, a high-level diffusion policy, and a low-level RL whole-body controller (Nai et al., 6 Feb 2026). The method achieved $5\,\mathrm{N}$3 success on unseen environments in a squatting task, $5\,\mathrm{N}$4 on kneeling and ring pickup, $5\,\mathrm{N}$5 on dynamic tossing, $5\,\mathrm{N}$6 on sword unsheathing, and $5\,\mathrm{N}$7 on walking to clean a table. Human-in-the-loop kinematic adaptation was important: removing it reduced marriage-proposal performance from $5\,\mathrm{N}$8 to $5\,\mathrm{N}$9.

World-action models and foundation-model coordination

WholeBodyWAM extends pretrained World Action Models with a distinct whole-body-control stream (Li et al., 15 Sep 2026). The model retains a pretrained $10\,\mathrm{N}$0-billion-parameter visual-manipulation pathway and adds a Unified Whole-Body Controller Interface with shared semantic fields for base velocity, pelvis height, torso orientation, lower-body positions, lower-body velocities, and controller-specific residuals.

Coordination-Aware Self-Attention increases manipulation-to-whole-body information flow when task-directional arm manipulability decreases. This allows the model to learn when stepping, pelvis motion, torso rotation, or body repositioning should support manipulation. The method achieved $10\,\mathrm{N}$1 overall simulation task success, $10\,\mathrm{N}$2 in-distribution real-world success, and $10\,\mathrm{N}$3 out-of-distribution real-world success. It reduced cross-controller success-rate variance to $10\,\mathrm{N}$4 percentage-point squared, compared with $10\,\mathrm{N}$5 for the cited baseline.

Other foundation-model systems use active spatial perception, occupancy-grid navigation, reachability optimization, foundation-model grasp synthesis, and manipulation primitives (Liang et al., 20 May 2026). Their architecture separates high-level spatial reasoning from geometric and dynamical execution. The reported hard-setting real-world success rates were $10\,\mathrm{N}$6, $10\,\mathrm{N}$7, $10\,\mathrm{N}$8, $10\,\mathrm{N}$9, and $\mu=0.7$0 across five task categories. The system illustrates a recurring design principle: large multimodal models handle semantic and relational reasoning, while classical planning, inverse kinematics, locomotion policies, and low-level control handle metric feasibility.

5. Non-prehensile, multi-contact, and soft whole-body manipulation

Quadrupedal body pushing

Learning Whole-body Manipulation for Quadrupedal Robot treats the entire quadruped as a non-prehensile manipulator (Jeon et al., 2023). The robot approaches an object, establishes body contact, and generates locomotion and contact motions that translate and rotate the object. A recurrent latent embedding captures object properties from action–observation history. The system does not use an explicit analytical object model during execution.

The method handles boxes and cylinders with randomized dimensions, mass, inertia, center of mass, friction, and drag. Simulation success was $\mu=0.7$1 under a $\mu=0.7$2, $\mu=0.7$3 criterion, and the real robot succeeded in all $\mu=0.7$4 reported trials. Its restrictions include planar pushing, reliance on pose sensing, restricted object classes, absence of obstacle avoidance, and no grasping or lifting.

Humanoid embracing

Humanoid embracing extends WBM beyond terminal grasping by using arms and torso to establish distributed contacts around bulky objects (Zheng et al., 16 Sep 2025). The system combines a human-motion prior, teacher–student distillation, reinforcement learning, and a neural signed distance field representing upper-body geometry. The task consists of approaching, embracing, and transporting an object.

In simulation, tested objects included cylinders, cuboids, and spheres. With NSDF, standard $\mu=0.7$5 objects generally achieved $\mu=0.7$6 success, while $\mu=0.7$7 objects achieved $\mu=0.7$8–$\mu=0.7$9 depending on shape. The real-world experiment used a Unitree H1-2 humanoid and a cylinder approximately $70\,\mathrm{Hz}$0 in diameter and $70\,\mathrm{Hz}$1 high. The paper reports successful whole-body embracing and transport, but not force measurements, maximum payload, or a direct comparison with end-effector-only control.

Soft whole-body manipulation

Zero-shot Whole-Body Manipulation with a Large-Scale Soft Robotic Torso demonstrates forceful lifting with two continuum soft arms and a vertically moving torso (Johnson et al., 28 Sep 2025). Baloo uses twelve bending configuration variables across two pneumatically actuated arms and a torso elevator. Each arm is simulated with chains of rigid disks and universal joints, while pneumatic chambers are represented through spatial tendons and pressure-state actuators.

The learned policy outputs thirteen reduced-dimensional pressure commands and elevator commands. A simple approach–grasp motion primitive guides RL exploration, whereas ordinary proximity, contact, and height reward shaping produced premature grasps and tipping. The final policy learns reactive behaviors, including re-grasping and perturbation recovery, but can also over-correct.

Simulation achieved $70\,\mathrm{Hz}$2 success on $70\,\mathrm{Hz}$3 randomized boxes. Hardware transfer without fine-tuning achieved $70\,\mathrm{Hz}$4 success on boxes between $70\,\mathrm{Hz}$5 and $70\,\mathrm{Hz}$6. The system does not use tactile or force sensing. Its transfer is attributed to passive compliance, approximate model calibration, broad object randomization, action and observation matching, and the relatively broad success margin of large-scale lifting.

Proprioceptive and blind manipulation

Blind Dexterity studies whole-body manipulation using only joint encoders, joint velocities, previous actions, projected gravity, and angular velocity (Bhatt et al., 30 Aug 2026). The central premise is that compliant position-controlled joints convert tracking deviations into an implicit whole-body contact signal. A contact changes joint positions, velocities, and tracking residuals, allowing the policy to infer object presence, resistance, motion, and pose from short histories.

Tasks include IMU-free push-resilient walking, football trapping, suitcase handle seeking and lifting, and skateboard seeking and mounting. In the football task, the policy learns to sweep and wiggle the foot to locate the ball. In the skateboard task, it probes the nose and tail, drags the board to infer yaw, and coordinates both feet for mounting. In the suitcase task, it touches the table, probes the suitcase sides, searches for the recessed handle, and lifts.

The policies use five-step histories for manipulation and are trained with PPO in Isaac Sim. Auxiliary estimators demonstrate that object state becomes more decodable after informative contacts, but estimator feedback is not consistently beneficial. The approach is limited by low spatial resolution, ambiguity, short memory, task-specific estimators, seed sensitivity, fragile-object interactions, and dependence on matching compliant actuation between simulation and hardware.

6. Evaluation, limitations, and research directions

Evaluation dimensions

WBM studies use heterogeneous metrics because the objectives differ across platforms and tasks. Reported measures include:

• end-effector position and orientation error;
• base velocity and yaw-rate error;
• reachable workspace;
• contact or stumble rate;
• planning iterations and planning time;
• success rate under object-pose tolerances;
• force-error and torque contributions;
• energy-tank lower-bound maintenance;
• human-applied force distributions;
• object transport and reorientation success;
• cross-controller variance;
• task progress before failure;
• sim-to-real success;
• action smoothness and trajectory continuity.

The metrics must be interpreted with respect to the control abstraction. A controller that reports an end-effector force command does not necessarily establish force tracking unless the force is measured. A planner that enforces robot-side contact constraints does not necessarily model object motion. A learned policy with high task success may not provide interpretable mass, friction, or contact estimates.

Common limitations

Several limitations recur across the literature.

Incomplete environmental dynamics: Door dynamics, human biomechanics, load inertia, object compliance, friction variation, and contact transitions are often omitted or treated implicitly. Receding-horizon correction, passive compliance, or learned adaptation can compensate experimentally, but these mechanisms do not replace a complete interaction model.

Restricted contact topology: Some optimization methods assume one point contact, planar motion, convex objects, fixed gait schedules, or predefined contact modes. General three-dimensional multi-contact manipulation remains substantially more difficult.

Computational cost: Full-order dynamics and contact optimization can be non-convex and expensive. Explicit surface optimization has improved planning time, but reported replanning can still take minutes. Learning-based policies provide fast inference but require simulation, demonstrations, or extensive policy training.

Perception and state estimation: Whole-body tasks involve viewpoint changes, occlusion, contact-induced motion, and body–object geometry. Systems based on motion capture or AprilTags provide reliable pose information but do not establish autonomous perception. Systems based on foundation models improve semantic flexibility but may remain weak at metric waypoint generation.

Safety and stability guarantees: Many RL systems rely on soft penalties, simulation, domain randomization, and low-level PD tracking rather than formal contact, balance, or stability constraints. Model-based approaches provide stronger constraint representations, but formal guarantees remain uncommon for the complete robot–object–environment system.

Embodiment dependence: Policies and interfaces often assume a particular morphology, actuation structure, camera arrangement, or WBC semantic interface. Cross-embodiment transfer is possible when a shared interface is maintained, but new robots generally require locomotion-policy training, calibration, or retargeting.

Data requirements: Demonstration-based methods reduce reward-engineering demands but require retargeting and hardware alignment. Robot-free systems increase collection efficiency, yet usually benefit from a small amount of direct robot data. Foundation-model systems reduce task-specific training requirements but depend on pretrained models, geometric modules, and downstream controllers.

Research directions

The supplied research indicates several converging directions:

1. Unified force, motion, and object reasoning: Future systems are likely to combine full-body inverse dynamics with object-level interaction models, contact uncertainty, and force sensing.
2. Continuous contact representations: Differentiable robot and object surfaces can replace dense contact sampling and support optimization over body-wide contact locations.
3. Hierarchical learned–model-based control: Learned policies can select posture, contact, and task intent, while MPC, inverse kinematics, barrier functions, or inverse dynamics enforce local feasibility.
4. Active perception through contact: Contact can serve both manipulation and sensing. Proprioceptive histories, compliant actuation, and deliberate probing can reveal object geometry and state under occlusion.
5. Terrain- and support-aware manipulation: Foot-contact planes, terrain encoders, and support-aware target representations can decouple manipulation objectives from locomotion-induced body fluctuations.
6. Semantic controller interfaces: Shared physical semantics for pelvis, torso, base, and lower-body commands can improve portability across heterogeneous whole-body controllers.
7. Human-compatible data collection: Wearable exoskeletons, VR interfaces, robot-free demonstrations, and human-in-the-loop kinematic adaptation can expand whole-body datasets while reducing robot hardware burden.
8. Safety-aware reactive learning: Re-grasping and perturbation recovery should be combined with tactile or force sensing, uncertainty estimation, contact constraints, and calibrated corrective actions.
9. Three-dimensional multi-contact WBM: Planar pushing and embracing provide useful testbeds, but general manipulation involving impacts, articulated objects, non-convex objects, multiple simultaneous contacts, and dynamic lifting remains unresolved.
10. General-purpose visuomotor whole-body policies: Pretrained world-action and diffusion models can preserve manipulation priors while adding explicit whole-body coordination, but their computational requirements, controller grounding, and real-time execution remain open problems.

WBM is therefore best understood not as a single controller architecture but as a research area organized around a systems principle: manipulation, locomotion, posture, contact, sensing, and balance should be treated as mutually coupled resources. Model-based MPC provides predictive dynamics and explicit constraints; whole-body inverse dynamics provides torque-consistent force and motion generation; contact optimization exposes body-wide contact choices; reinforcement learning supplies adaptation and complex behavior; teleoperation and motion retargeting provide scalable demonstrations; and proprioceptive or exteroceptive sensing supports interaction-aware control. The principal unresolved challenge is to combine these capabilities into systems that are simultaneously general, physically grounded, computationally tractable, perceptually robust, and formally safe.