Learning Force Control for Legged Manipulation

The Big Picture

Imagine trying to thread a needle while wearing oven mitts. That’s roughly the challenge facing most robot arms when they interact with the physical world: they know where to move, but they have no intuitive sense of how hard to push. Too little force and the task fails. Too much, and something breaks.

For four-legged robots like quadrupeds, the problem gets worse. Every step the legs take ripples through the entire body and changes the forces at the arm’s tip.

Researchers at MIT’s Improbable AI Lab took a different approach. Most robot controllers today produce force only as a side effect of position commands. They keep moving the arm until it reaches a target location, and force is whatever happens along the way. What this team built instead is a controller that can be directly told how hard to push, coordinating all four legs and the arm together to hit that target precisely.

The result: a four-legged robot with an arm that has calibrated, controllable touch.

Key Insight: By training a reinforcement learning policy to track explicit force commands, without any dedicated force sensors, the researchers achieved compliant whole-body manipulation on a quadruped robot. Teleoperation becomes more intuitive, and the approach brings legged robots closer to working safely alongside people.

How It Works

The team reframed the learning problem. Standard sim-to-real reinforcement learning (RL) trains a robot policy in a physics simulation, then deploys it in the real world. This dominates legged locomotion research, but those policies typically learn to achieve positions, not forces. The MIT team added a force-tracking objective directly into the training loop.

Rather than equipping the robot with expensive wrist force-torque sensors, the policy learns to estimate contact forces entirely from proprioception (the robot’s internal sense of its own joint positions, velocities, and motor torques). No extra hardware. The policy infers how hard the gripper is pushing from the same signals a human feels as muscle tension and joint resistance.

Training happens in a carefully designed simulation environment. The researchers place the robot in a virtual “potential field,” a simulated spring pulling the gripper toward a randomized target point. This forces the policy to push against varying resistances and learn what different force magnitudes feel like proprioceptively. The policy receives three types of commands simultaneously:

• A base velocity command telling the body where to walk
• An end-effector position command telling the arm where to place the gripper (the end-effector is the tool at the tip of the arm)
• A force command specifying how hard to press at the gripper

A single neural network, the whole-body controller, coordinates all four legs and the arm at once. When the arm needs to push harder, the legs automatically shift weight and reposition the torso to provide a stable foundation. The human teleoperator doesn’t manage any of this. They command “push with 10 Newtons” and the robot handles the rest.

This setup recovers two classical control behaviors roboticists have long wanted from learned systems. Gravity compensation makes the arm feel weightless when a human physically guides it; the robot cancels out its own arm’s weight so a person can move it through tasks by hand. Impedance control lets the operator set the gripper’s effective stiffness: high for precise positioning, low for compliant contact with fragile or irregular surfaces. Both come from the same learned force controller, just with different commanded force profiles.

Why It Matters

Legged robots have gotten good at locomotion: running, jumping, navigating rubble, recovering from falls. Manipulation robots can do dexterous grasping and assembly. But combining both with enough finesse to interact safely around humans and fragile environments has stayed stubbornly difficult.

Force control is what’s missing for a wide class of real-world tasks. Think of a robot that needs to wipe a surface without scratching it, tighten a bolt to a specific torque, or help an elderly person stand without jerking. All of these require knowing not just where to move, but how hard to push at every instant.

That the MIT team pulled this off without force sensors matters for deployment. Sensors add cost, fragility, and failure modes. A robot that infers forces from its own proprioception is cheaper and more durable.

The teleoperation angle matters too. Because force is now a first-class commanded quantity, a single human operator can demonstrate complex manipulation tasks: pulling open drawers, handling objects with varying weights, applying sustained pressure. That demonstration data can feed future imitation learning systems, potentially letting robots learn manipulation skills from watching people work.

Bottom Line: MIT researchers trained a quadruped robot to directly control the forces it applies, without force sensors, by learning from simulation alone. This is the first legged robot to deploy learned whole-body force control in the real world, enabling compliant manipulation that could reshape how robots work alongside humans in unstructured environments.