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

TL;DR

The paper tackles generalization challenges in loco-manipulation by introducing a contact-explicit, unified task representation that treats contact as a fundamental primitive. A high-level planner outputs sequences of contact goals, and a single goal-conditioned RL policy learns to realize these plans across diverse morphologies and tasks. Empirical results on a quadruped, a humanoid, and a humanoid performing bi-manual manipulation show improved generalization to unseen velocities, directions, and object shapes, with the multi-task policy outperforming task-specific or non-contact baselines. The approach suggests a scalable foundation for loco-manipulation with strong sim-to-real potential via domain randomization and dense contact-focused rewards.

Abstract

We present a unified framework for multi-task locomotion and manipulation policy learning grounded in a contact-explicit representation. Instead of designing different policies for different tasks, our approach unifies the definition of a task through a sequence of contact goals-desired contact positions, timings, and active end-effectors. This enables leveraging the shared structure across diverse contact-rich tasks, leading to a single policy that can perform a wide range of tasks. In particular, we train a goal-conditioned reinforcement learning (RL) policy to realise given contact plans. We validate our framework on multiple robotic embodiments and tasks: a quadruped performing multiple gaits, a humanoid performing multiple biped and quadrupedal gaits, and a humanoid executing different bimanual object manipulation tasks. Each of these scenarios is controlled by a single policy trained to execute different tasks grounded in contacts, demonstrating versatile and robust behaviours across morphologically distinct systems. Our results show that explicit contact reasoning significantly improves generalisation to unseen scenarios, positioning contact-explicit policy learning as a promising foundation for scalable loco-manipulation.

Figures (7)

• Figure 1: Snapshots of our contact-explicit framework in action. (Row 1): the quadruped demonstrates diverse gaits; (Row 2): the humanoid demonstrates locomotion using hand-assisted gaits such as quadrupedal jump, pace and bipedal gaits such as walk, jump; (Row 3): the humanoid carries out different bimanual manipulation tasks, where the green box represents the target pose.
• Figure 2: Overview of our contact-explicit framework. A high-level planner generates the contact goals (and object pose targets for manipulation), that is provided as immediate goals for the goal-conditioned RL policy to accomplish.
• Figure 3: Simple illustration of a robot's end effector during different phases of contact, for a fixed command duration $S$. During the detach phase, the end effector is detached from a contact and is free to move. During the reach phase, the end effector is guided towards the desired contact location. During the hold phase, it maintains the contact.
• Figure 4: Quadruped robot crossing a gap with a bound gait by accurately adjusting the contact locations to remain on the wooden terrain.
• Figure 5: Comparison of velocity tracking error in all $x$-$y$ directions. Each cell in the grid is a combination of $x$ and $y$ velocity. The red line denotes the velocity combinations seen during training. The results were averaged over 500 episodes (each lasting 15 seconds of simulation time).