Loco-Manipulation

• Also called: loco manipulation, locomanipulation, whole-body loco-manipulation, legged mobile manipulation
• Category: Control › Locomotion control
• Tags: humanoid, legged-robots, whole-body-control, manipulation, locomotion, mobile-manipulation

The coordinated use of a robot's locomotion and manipulation abilities, where moving the body and handling objects are planned and controlled together rather than as separate steps.

Definition

Loco-manipulation refers to tasks in which a robot must move its body through the environment and physically interact with objects at the same time, or in a tightly coupled sequence. Examples include a humanoid carrying a box up a ramp, a quadruped with an arm opening a door and stepping through it, or a legged robot pushing a cart. The term joins locomotion and manipulation to stress that the two cannot be treated independently when the robot's base is mobile and its contacts with the world keep changing.

Classical mobile manipulation often decouples the problem: drive the base to a location, stop, then use the arm. Loco-manipulation instead treats the whole body as one system. Forces applied by the hands or end effector affect balance, foot placement and base motion, and the base's motion in turn changes what the arms can reach. A humanoid lifting a heavy object must shift its center of mass; a quadruped pressing a button with a front leg loses that leg as a support.

Approaches fall into two broad families. Model-based methods use whole-body control and optimization, such as model predictive control over contact sequences and centroidal dynamics, solving tasks for the feet, hands and torso together under friction and torque limits. Learning-based methods train policies with reinforcement learning in simulation or learn from human motion data, often retargeting human demonstrations to the robot's body. Hybrid systems combine a learned locomotion policy with planners or model-based controllers for the manipulation targets.

Key difficulties include planning sequences of contacts that involve both the environment and the manipulated object, regulating interaction forces while staying balanced, and obtaining enough training data that captures realistic human–object interaction. Platforms used for this work include humanoid robots, quadrupeds fitted with arms, and wheeled-legged robots.

Why it matters

Many useful tasks in homes, warehouses and disaster sites require a robot to carry, push, open or climb while moving, not just to walk or just to grasp. Loco-manipulation is the capability that turns legged and humanoid robots from mobile platforms into machines that can do physical work in spaces built for people.

Key applications

• Humanoid robots carrying, lifting and placing objects while walking
• Quadrupeds with arms opening doors, pressing buttons and retrieving items
• Pushing and pulling carts, trolleys and heavy objects
• Climbing onto and interacting with furniture, stairs and other scene elements
• Inspection and intervention tasks at industrial or disaster sites

In practice

A team working on a humanoid, or on a quadruped with an arm, usually starts from a reliable locomotion controller and then adds manipulation objectives to it, either as extra tasks in a whole-body controller or as additional commands and rewards for a learned policy. In simulation they randomize object mass, friction and contact positions so the policy learns to stay balanced under unexpected loads. Human motion capture or teleoperation data is often retargeted to the robot to provide reference motions for interactions such as picking up boxes or sitting on furniture. On hardware, engineers test with a range of payloads and monitor joint torques and foot contact states to see where the controller trades balance against task performance.

Common misconceptions

• Loco-manipulation is just a mobile base with an arm bolted on — the defining feature is that body motion and interaction forces are coupled and must be controlled together
• Good locomotion and good manipulation controllers can simply be combined — forces from manipulation disturb balance and require coordinated whole-body control or joint training
• It only applies to humanoids — quadrupeds with arms, wheeled-legged robots and robots that manipulate with their legs are all studied under the same term

Who should know this

• Humanoid and legged robotics researchers
• Control engineers working on whole-body control
• Robot learning teams training policies for mobile manipulation
• Developers integrating arms onto quadruped platforms

Related technologies

• Whole-Body Control
• Model Predictive Control
• Reinforcement Learning
• Motion Retargeting
• Centroidal Dynamics
• Contact-Implicit Optimization

Related terms

• Whole-Body Control
• Locomotion
• Humanoid Robot
• Quadruped
• Dexterous Manipulation

Related projects

• LeRobot: HuggingFace ML library for real-world robotics in PyTorch
• Isaac Lab: NVIDIA robot learning framework: one task API across Newton, PhysX and Isaac Sim
• NVIDIA Isaac GR00T: NVIDIA open foundation model for humanoid robot reasoning and skills
• OpenVLA: 7B-parameter Vision-Language-Action model for robotic manipulation
• Open Duck Mini V2: An open-source mini duck robot you 3D-print and build, inspired by Disney's BDX droid
• Mobile ALOHA: Bimanual mobile manipulation with 4 robot arms on a wheeled base for teleoperation