Skip to content

HumanoidSoccerMaze

A novel benchmark requiring a humanoid robot to navigate a maze while dribbling a ball to a distant target, with a baseline solution that integrates planning and learned locomotion.

What this is

HumanoidSoccerMaze tests a Booster T1 humanoid on goal-conditioned maze navigation combined with ball dribbling. The robot must find a path through a maze with walls and obstacles, and execute it using a learned locomotion-dribbling policy.

The benchmark is built on mjlab (GPU-accelerated RL via MuJoCo Warp) and uses the Booster T1 23-DOF humanoid — the platform used in the RoboCup Humanoid Soccer League.

An example of evaluation after approximately 200 million steps of training on a RTX 4090 in approximately 3 hours:

The planning–learning interface

The core challenge is that a low-level dribbling policy cannot reason about walls several meters away, while a planner cannot control 23 joints to keep the ball moving smoothly. The solution couples them through a shared direction signal.

┌─────────────────┐          direction vector         ┌──────────────────────┐
│    Planner      │  ─────────────────────────────>   │  Dribbling Policy    │
│  (discrete,     │                                   │  (continuous,        │
│   global)       │  <─────────────────────────────   │   local)             │
│                 │     current robot/ball position   │                      │
└─────────────────┘                                   └──────────────────────┘

Planner side — a classical planner solves the maze using a Sokoban abstraction: it models the ball-dribbling problem as a discrete push-planning problem on a grid. Given the current ball cell and the goal cell, it computes a feasible path through free cells considering the Sokoban constraints ( the ball can only be pushed, not pulled). The planner outputs a direction vector pointing from the robot's current cell toward the next cell in the plan.

Learning side — the dribbling policy (trained as in the dribbling task) observes a local direction vector pointing toward the next waypoint in the plan. It uses this to steer its dribbling (or just walking).

AbstractionManager bridge — at each environment step, the GridAbstractionTerm re-reads the robot position and ball position, maps them to grid cells, looks up the pre-computed cost map, and provides a direction_at_pos() signal to observation functions. The planner only replans when the goal cell or map layout changes (change-detection via AbstractionSettings), avoiding expensive replanning every step.

env step:
  abstraction_manager.compute(dt)
    └── GridAbstractionTerm._update_signals()
          └── look up direction_map at current robot cell
                → feeds "maze_direction" observation term

The maze layout and goal are provided as part of the environment config and can be randomised across episodes for generalisation.

Task structure

The environment extends the standard dribbling task:

  • Same robot and ball setup as t1-dribbling.
  • Additional scene elements: maze walls encoded as a binary occupancy grid.
  • Additional observation term: maze_direction — a 2D unit vector in robot body frame pointing toward the next grid waypoint.
  • Same reward structure: ball velocity tracking + locomotion regularization. The planner direction acts as a soft constraint through the observation, not as a separate reward term.

Code structure

src/colosseum/tasks/soccer_maze/     # Benchmark environment (scene, MDP, rewards)

src/colosseum/mdp/abstraction/maze/
└── grid_abstraction.py              # GridAbstractionTerm — planner + signal cache

Quick start

# Train
pixi run train task:t1-soccer-maze

# Evaluate
pixi run play task:t1-soccer-maze --checkpoint ./logs/<run>/checkpoints/latest.pt

Paper

The paper describes the technical contribution of the benchmark (environment features, performance metrics) and the baseline solution approach. It is under submission at RoboCup Symposium 2026.

HumanoidSoccerMaze: a novel benchmark and a hierarchical solution approach
Flavio Maiorana, Daniel Gigliotti, Luca Iocchi
RoboCup Symposium 2026 (under submission)