💻 computer science

Learning to Act Through Contact: A Unified View of Multi-Task Robot Learning

This paper introduces a unified framework that employs a single goal-conditioned reinforcement learning policy to master diverse locomotion and manipulation tasks across different robotic morphologies by explicitly defining and executing sequences of contact goals.

Original authors: Shafeef Omar, Majid Khadiv

Published 2026-05-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shafeef Omar, Majid Khadiv

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to move. Traditionally, if you wanted a robot to walk, you'd teach it "walk forward." If you wanted it to pick up a cup, you'd teach it "grab cup." If you wanted it to jump, you'd teach it "jump." The robot learns these as separate, isolated tricks. If you ask it to do something slightly new, like walking on stepping stones instead of flat ground, it often freezes because it hasn't been taught that specific trick.

This paper proposes a different way of thinking. Instead of teaching the robot what to do (walk, jump, grab), the authors teach the robot where to touch.

The Core Idea: The "Touch List"

Think of the robot's brain not as a list of commands like "walk" or "dance," but as a schedule of handshakes.

In this new framework, a "task" is just a sequence of contact goals.

Goal 1: "Touch the ground with your left foot at this specific spot for 0.5 seconds."
Goal 2: "Touch the ground with your right foot at that spot for 0.5 seconds."
Goal 3: "Touch the table with your hand here."

The robot doesn't care if it's walking, crawling, or lifting a box. It only cares about the schedule of touches. If the schedule says "touch here, then touch there," the robot figures out the best way to move its body to make those touches happen.

The Three "Modes" of Touching

The authors break down every interaction into three simple phases, like a dance routine:

Reach: The robot swings its limb out to find the target spot (like a hand reaching for a doorknob).
Hold: Once it touches, it stays there firmly (like holding the doorknob).
Detach: It lets go and moves freely to find the next spot (like letting go to walk away).

By mixing and matching these three phases, the robot can do almost anything.

The "Swiss Army Knife" Robot

The researchers tested this idea on two very different types of robots: a four-legged dog (quadruped) and a two-legged human-like robot (humanoid). They trained one single brain (one policy) to handle everything.

The Dog: This single brain learned to trot, pace, bound, jump, and even crawl. It didn't need a different brain for each gait. It just followed the "touch schedule."
The Humanoid: This robot learned to walk on two legs, crawl on four, and even do a "hand-assisted" jump.
The Hands: The same humanoid brain learned to pick up a box, rotate it on a table, and lift it while keeping track of its position.

Why This is a Big Deal (The Analogy)

Imagine you are learning to play the piano.

Old Way: You memorize a specific song. If someone asks you to play a different song, you can't. You have to relearn everything from scratch.
New Way (This Paper): You learn the concept of "pressing keys at specific times." You don't memorize songs; you memorize the rhythm and timing of the notes. Because you understand the underlying logic of "when to press," you can play a song you've never heard before just by reading the sheet music (the contact schedule).

The paper shows that by focusing on contact (the "when and where to press"), the robot becomes much better at adapting to new situations.

Real-World Proof

The researchers showed that this approach works better than the old methods in two key ways:

New Directions: When asked to walk sideways (something it wasn't explicitly trained to do), the "touch-based" robot figured it out immediately. The old "velocity-based" robot just stood there confused.
New Shapes: When asked to manipulate a round ball instead of a square box (shapes it had never seen), the "touch-based" robot adapted its grip instantly. The old robot, which relied on memorizing "box shapes," struggled.

The Bottom Line

The paper claims that by treating touch as the fundamental building block of movement, rather than just a side effect of moving, we can create one universal robot brain. This brain can switch between walking, jumping, and lifting objects seamlessly, simply by following a new list of "where to touch" instructions. It makes robots more flexible, robust, and ready for the messy, unpredictable real world.

Technical Summary: Learning to Act Through Contact

Problem Statement
Current reinforcement learning (RL) approaches for robotics often rely on task-specific objectives, resulting in policies that struggle to generalize to unseen scenarios without retraining. While RL has achieved success in specific domains like agile quadruped locomotion and dexterous manipulation, prevailing methods typically treat contact as an incidental byproduct of motion optimization rather than a fundamental control primitive. This abstraction limits the transferability of skills across diverse physical interactions (e.g., from walking to object manipulation) and across different robot morphologies. Existing goal representations, such as 3D position targets or motion references, often overlook the critical importance of contact timing and placement, making it difficult to unify locomotion and manipulation under a single framework.

Methodology
The authors propose a unified framework grounded in a contact-explicit representation. Instead of designing separate policies for different tasks, the approach defines tasks through sequences of contact goals. These goals specify:

Target contact positions ( $p^{con}$ )
Contact timings/durations ( $S$ )
Active end-effectors (binary indicators $I^{con}$ )
Object pose targets (for manipulation tasks)

The system operates via a high-level planner that generates these contact goal sequences, which are then executed by a goal-conditioned RL policy. The policy is trained to track these goals using joint torques, regardless of the specific robot morphology or task type.

Contact Phases and Rewards
The framework decomposes interaction into three distinct phases for each end-effector, enabling controlled contact creation and breaking:

Reach (R): The end-effector moves toward a desired contact location.
Hold (H): The end-effector maintains contact at the achieved location.
Detach (D): The end-effector moves freely without engaging contact.

To train the policy, the authors employ a dense reward structure that incentivizes specific behaviors based on the current phase:

Reach Reward: Encourages minimizing the distance to the target contact location when the contact indicator is 0.
Hold Reward: Incentivizes maintaining contact at the desired location when the indicator is 1.
Detach Reward: Penalizes unintended contact when the indicator is 0.
Pose Reward: For manipulation, an additional term tracks the object's position and orientation.

The policy observes proprioceptive states (joint positions/velocities), the current and next contact goals, and command durations. Notably, the framework does not rely on explicit contact sensing from the robot's feet for training in simulation.

Key Contributions

Unified Contact-Explicit Framework: The authors propose a goal-conditioned RL framework where tasks are defined by contact goals, enabling a single policy to learn diverse locomotion gaits and manipulation skills.
Empirical Validation Across Morphologies: The framework is validated on morphologically distinct robots (quadrupeds and humanoids) performing a wide range of tasks, demonstrating that explicit contact reasoning facilitates robust, dynamic behaviors across different embodiments.

Experimental Results
The framework was evaluated on three primary scenarios:

Quadruped Locomotion: A single policy successfully executed multiple gaits (trot, pace, bound, jump, crawl) and navigated challenging terrains, including gap crossings.
Humanoid Locomotion: A single humanoid policy performed both bipedal (walk, jump) and quadrupedal (crawl, pace, jump) gaits, including hand-assisted movements.
Bimanual Manipulation: A humanoid robot performed complex tasks such as reorienting objects on a table and tracking object poses while lifting them.

Generalization Findings:

Unseen Velocities/Directions: Compared to policies trained on velocity targets, the contact-explicit policy generalized significantly better to lateral movements and out-of-distribution velocity directions, despite not receiving specific rewards for lateral motion during training.
Multi-Gait vs. Single-Gait: The multi-gait policy demonstrated lower contact plan deviation (Hamming distance) across varying command durations compared to single-gait policies, suggesting effective interpolation between gaits.
Unseen Object Shapes: In manipulation tasks, the contact-explicit policy generalized better to unseen object shapes (cylindrical and spherical) compared to a baseline using one-hot task encoding. The contact-explicit policy exhibited emergent behaviors, such as retrying strategies, to track poses on rolling spherical objects.
Unseen Object Poses: The contact-explicit policy achieved lower tracking errors and variability on object poses extrapolated beyond the training distribution compared to the one-hot baseline.

Significance and Claims
The paper claims that treating contact as a central physical primitive, rather than a byproduct of optimization, allows for a more fundamental abstraction of robot tasks. By unifying locomotion and manipulation through contact goals, the approach enables a single policy to generalize across morphologically distinct systems and unseen scenarios. The authors position contact-explicit policy learning as a promising foundation for scalable loco-manipulation, arguing that it leverages shared structures across tasks to improve adaptability without the need for retraining from scratch for every new task. Future work is proposed to integrate this low-level policy with learned high-level planners for autonomous long-horizon tasks and to explore sim-to-real robustness.