Humanoid Goalkeeper: Learning from Position Conditioned Task-Motion Constraints

TL;DR

The paper addresses the challenge of humanoid goalkeeping by proposing an end-to-end RL policy that jointly optimizes task performance and motion realism through region-conditioned, adversarial motion priors. By conditioning on ball state and partitioning the interception space into regions $\mathcal{R}$, the method learns region-specific, human-like whole-body motions and uses a ball estimator to bridge perception gaps, enabling robust sim-to-real transfer with either MoCap or onboard perception. Hardware experiments demonstrate agile, autonomous interceptions across a wide goal area and reveal improved generalization to related dynamic tasks (escaping, grabbing) while highlighting perceptual and planning limitations that guide future work. The approach advances lifelike humanoid-object interactions and offers a scalable framework for integrating perception, decision-making, and agile motor control in real-world settings.

Abstract

We present a reinforcement learning framework for autonomous goalkeeping with humanoid robots in real-world scenarios. While prior work has demonstrated similar capabilities on quadrupedal platforms, humanoid goalkeeping introduces two critical challenges: (1) generating natural, human-like whole-body motions, and (2) covering a wider guarding range with an equivalent response time. Unlike existing approaches that rely on separate teleoperation or fixed motion tracking for whole-body control, our method learns a single end-to-end RL policy, enabling fully autonomous, highly dynamic, and human-like robot-object interactions. To achieve this, we integrate multiple human motion priors conditioned on perceptual inputs into the RL training via an adversarial scheme. We demonstrate the effectiveness of our method through real-world experiments, where the humanoid robot successfully performs agile, autonomous, and naturalistic interceptions of fast-moving balls. In addition to goalkeeping, we demonstrate the generalization of our approach through tasks such as ball escaping and grabbing. Our work presents a practical and scalable solution for enabling highly dynamic interactions between robots and moving objects, advancing the field toward more adaptive and lifelike robotic behaviors.

Figures (9)

• Figure 1: Method framework: We train our policy using an end-to-end reinforcement learning (RL) framework, incorporating position-conditioned task and motion constraints. Our method is capable of operating with either an onboard camera or the motion-capture (MoCap) system.
• Figure 2: This design ensures that when the ball is far from the robot, the interaction is based on the predicted landing point, whereas for closer balls, the target position is the actual ball location, allowing for more precise handling near the robot.
• Figure 3: The motion curation and formulation pipeline, demonstrating the process from human video to robot-executed motion.
• Figure 4: Each scatter point represents a trial, with the position corresponding to the ball target of the trial. The color gradient of the points indicates $E_{\text{pos}}(\text{min})$ for each trial. The pentagram represents the cluster center of trials that has the same matching region (regions with minimal $E_{\text{pos}})$. A well-learned position-conditioned policy should exhibit a clear correspondence between the cluster centers and their respective ball target regions
• Figure 5: We visualize the 3D trajectory of the ball, as well as the left and right hand trajectories for one trial in each region, and compare the performance of the proposed method with that of the non-division baselines.