💻 computer science

Vision-Based Dribbling for Humanoid Soccer via Privileged Representation Learning

This paper proposes an integrated reinforcement learning framework that embeds a temporal depth encoder into a policy to enable a humanoid robot to learn robust, vision-based soccer dribbling against static and dynamic opponents directly from onboard depth observations, achieving high success rates without relying on explicit state estimation or privileged information.

Original authors: Flavio Maiorana, Valerio Spagnoli, Eugenio Bugli, Flavio Volpi, Daniele Affinita, Vincenzo Suriani, Daniele Nardi, Luca Iocchi

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Flavio Maiorana, Valerio Spagnoli, Eugenio Bugli, Flavio Volpi, Daniele Affinita, Vincenzo Suriani, Daniele Nardi, Luca Iocchi

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 a robot soccer player trying to dribble a ball down a field while wearing a pair of high-tech goggles. Now, imagine that instead of having a supercomputer in its head that knows the exact speed and position of the ball and its enemies, this robot has to figure everything out just by looking at a video feed from its own eyes, all while trying not to fall over. That is the wild challenge tackled in this paper.

The researchers built a system where a humanoid robot learns to dribble a soccer ball by combining "seeing" and "moving" into one big brain. Usually, robots are built like a relay race: one team detects the ball, passes the info to a second team that plans the path, and a third team moves the legs. But the authors argue this old-school method is like trying to drive a car while only looking at a map drawn by someone else; if the map is slightly wrong or the ball gets hidden for a split second, the driver crashes. Instead, they taught the robot to learn perception and control together, like a human learning to ride a bike by feeling the balance rather than calculating the physics of every wobble.

To teach this robot, they used a clever two-step training camp. First, they let the robot practice in a video game simulation where it had "cheat codes" (called privileged information). It knew exactly where the ball and any enemies were, even if they were behind a wall. The robot learned to dribble perfectly in this cheat-mode, mastering how to keep the ball close and dodge obstacles. Then, in the second phase, they took away the cheat codes. They trained a special "vision encoder"—think of it as a student tutor—to look at the same depth-camera video the robot would see in the real world and guess what the "cheat code" numbers would have been. The robot's brain then used these guesses to make decisions, effectively learning to play the game using only its eyes.

The results of this training were impressive, but with some very specific limits. When the robot played on an empty field with no one trying to steal the ball, it was a perfect star, succeeding 100% of the time. When they added a single, stationary obstacle (like a cone or a wall) in its path, it still did a fantastic job, succeeding 96% of the time and only taking a tiny bit longer to reach the goal.

However, the story changes when the robot faces a moving enemy. When a second robot was programmed to chase the ball and try to knock it away, the success rate dropped to 46%. In these simulations, the robot fell over or lost the ball much more often, and it bumped into the opponent 68% of the time. The authors suggest that while the robot is great at seeing and moving on its own, the real challenge is reacting to a live, moving opponent that is actively trying to interfere. The robot's eyes were working fine—it could still see the ball and the enemy reasonably well—but the "brain" needed to figure out how to dodge a moving target without falling over is still a work in progress.

It is important to remember that all of this happened inside a computer simulation using a specific robot model called the Booster T1. The authors are careful to say this is a strong foundation for the future, but they haven't tested it on a real robot on a real field yet. They propose that this method of teaching robots to learn from video feeds while balancing is a promising path forward, but for now, the robot is still a very talented student in a virtual classroom, not quite ready for the World Cup.

Technical Summary: Vision-Based Dribbling for Humanoid Soccer via Privileged Representation Learning

Problem Statement
Humanoid soccer dribbling represents a complex loco-manipulation challenge requiring the simultaneous maintenance of dynamic balance, precise ball control, and awareness of a dynamic adversary under onboard sensing constraints. Existing approaches typically decouple perception from control, relying on modular pipelines (e.g., YOLO detection followed by Kalman filtering) to estimate ball and opponent states. While effective in controlled settings, this separation fails under occlusions, rapid ball movements, and complex opponent interactions because the perception module is not optimized for the downstream control policy. Furthermore, modern humanoid platforms (e.g., Booster T1) operate at higher speeds, increasing the demands on perception and real-time control. The specific challenge addressed is enabling a humanoid robot to dribble a soccer ball toward a target while avoiding static and dynamic obstacles using only onboard depth observations, without explicit state estimation or privileged scene information.

Methodology
The authors propose an integrated framework that learns perception and control jointly through a temporal encoder shaped by the policy's needs. The approach is inspired by Rapid Motor Adaptation (RMA) and utilizes a two-phase training scheme within a Partially Observable Markov Decision Process (POMDP) formulation:

Phase 1: Privileged Representation Learning (Curriculum Learning):
- Goal: Train a policy using "privileged" ground-truth state information (ball and opponent states) to establish a robust baseline for loco-manipulation.
- Curriculum: The training progresses through four stages: (0) ball approach with no obstacles; (1) target-directed dribbling with the ball near the robot; (2) dribbling with a static blocker; and (3) dribbling with a dynamic, ball-attacking opponent.
- Optimization: Stages 1–3 utilize a DAgger-regularized PPO objective, using the Stage 0 policy as a teacher to mitigate catastrophic forgetting. The reward function includes terms for ball velocity tracking, proximity, alignment, target progress, and penalties for collisions, falls, and unsafe contacts.
- Architecture: The policy receives proprioceptive inputs (IMU, joint states, gait phase) concatenated with a latent vector $z_t$ derived from the privileged state via a small MLP.
Phase 2: Visual Adaptation (Privileged Distillation):
- Goal: Replace the privileged encoder with a deployable visual encoder that infers the same latent representation $z_t$ from onboard depth observations.
- Architecture: A temporal depth encoder consisting of a Convolutional Neural Network (CNN) for spatial feature extraction and a Gated Recurrent Unit (GRU) for temporal aggregation.
- Training: The policy from Phase 1 is frozen. The visual encoder is trained to minimize the L2 distance between its output latent and the privileged latent ( $L_{latent}$ ). Auxiliary heads regress ball and obstacle states to improve identifiability ( $L_{adapt}$ ).
- Deployment: At inference, the robot uses only proprioception and the depth-based latent vector, with no access to ground-truth states.

Key Contributions

Temporal Encoder via Privileged Distillation: A method to infer task-relevant ball and opponent states directly from onboard depth observations, enabling deployable perception without explicit state estimation.
Obstacle-Aware Training Pipeline: A curriculum that progressively exposes the robot to increasing difficulty, from nominal control to static blockers and dynamic opponent-like obstacles.
End-to-End RL System: A unified policy framework integrating perception, balance control, ball handling, and opponent avoidance, released as open-source code.

Experimental Results
Experiments were conducted in simulation using the Booster T1 humanoid platform and the mjlab framework. The evaluation covered three conditions: no obstacles, a single static obstacle, and a single dynamic "ball-attacker" opponent.

Success Rates: The final policy achieved 100% success in the no-obstacle condition and 96% success with a static obstacle. Success dropped to 46% against an actively moving opponent.
Failure Modes: In the dynamic opponent scenario, the primary failure modes were falls (52% failure rate) and collisions (68% collision rate), indicating that robust closed-loop control against moving adversaries remains challenging.
Perception and Control: Velocity tracking errors remained stable between the no-obstacle and static obstacle conditions. In the dynamic scenario, larger angular and vector errors occurred during "blocked" timesteps (when the opponent was between the ball and target), suggesting the policy intentionally deviated from commanded ball motion to avoid collisions. Perception metrics (ball/obstacle position and velocity errors) remained relatively stable across conditions, indicating that the performance drop in the dynamic scenario was due to control difficulty rather than perception failure.
Ablation: Removing curriculum stages significantly degraded performance, particularly in obstacle scenarios, validating the necessity of progressive difficulty exposure.

Significance and Claims
The paper claims that the proposed framework supports robust vision-based dribbling in nominal and moderately dynamic settings. The results demonstrate that it is possible to learn opponent-aware dribbling directly from depth observations without explicit state estimation. However, the authors are modest regarding the dynamic scenario, acknowledging that robust dribbling against an actively moving adversary is still an open problem. The work establishes a foundation for handling more challenging moving-adversary scenarios and outlines a principled roadmap for sim-to-real transfer on the Booster T1 platform. The authors note that current evidence is limited by the simulation environment and the number of independently trained policies.

More like this