rs-imitation-learning-data

Game logs for training a network that plays RoboCup Small Size League, Division B (6 v 6) in the grSim simulator by imitating Sumatra, the TIGERs Mannheim AI.

Code, data format and training pipeline: https://gitlab.com/yogyanp/rs-imitation-learning.

Contents

Folder	Files	Size	What it is
selfplay/	42 games	3.4 GB	Sumatra vs Sumatra in grSim. Every robot of both teams has an expert action in every frame. About 13 h of play, 8 h of it running play.
selfplay/collect_stats.txt			one JSON line per game: duration, frames, why the collector restarted play
dagger/2026-10-10/r00 … r10	44 matches	1.2 GB	DAgger rounds: our network drove blue, while Sumatra ran in shadow mode and logged what it would have done in each blue robot's real state. Round rNN was played by the network trained after round N-1.

Each file is one game: a gzipped JSON-lines log (*.jsonl.gz), one object per world frame (about 60 per second). It holds the timestamp, game state, ball, all robots, and an actions list with Sumatra's command per robot: planned velocity now and 100 ms / 200 ms ahead, plus kick speed, chip, armed and dribble. Positions are in mm and velocities in m/s, in Sumatra's absolute field frame. The full field list is in shared_schema/SCHEMA.md §1.

A frame (numbers shortened; ... marks the other 11 robots and actions):

{"ts": 65184656999, "frame": 3003, "game_state": "RUNNING", "game_state_for": "NEUTRAL",
 "neg_half_team": "BLUE",
 "ball": {"x": 3981.7, "y": 2590.7, "z": 0.0, "vx": 0.249, "vy": 0.451, "vz": 0.0},
 "robots": [{"id": 1, "color": "BLUE", "x": -2709.9, "y": 699.7, "theta": 0.299,
             "vx": 0.064, "vy": -0.242, "w": 0.0, "ball_contact": false}, ...],
 "actions": [{"id": 1, "color": "BLUE", "cmd_vx": 0.050, "cmd_vy": -0.193, "cmd_w": 0.0,
              "cmd_vx_100ms": 0.027, "cmd_vy_100ms": -0.085, "cmd_w_100ms": 0.0,
              "cmd_vx_200ms": 0.0, "cmd_vy_200ms": 0.0, "cmd_w_200ms": 0.0,
              "kick_speed": 0.0, "is_chip": true, "is_armed": false, "dribble_speed": 0.0}, ...]}

Note that cmd_vx/vy/w is the plan at t = 0, which equals the robot's current velocity. The expert's intent is in the _100ms / _200ms fields, which training uses as the label.

Using it

git clone https://gitlab.com/yogyanp/rs-imitation-learning.git && cd rs-imitation-learning
bash scripts/download_data.sh                       # -> rl-framework/data/{selfplay,dagger}
cd training
python build_dataset.py --logs ../rl-framework/data/selfplay --out data/selfplay
python train.py --data data/selfplay --out checkpoints/runs/my_run

build_dataset.py reads the .jsonl.gz files directly. It turns each frame into one training sample per robot: a 135-number observation from that robot's team's point of view, plus the expert action. To download without the repo:

from huggingface_hub import snapshot_download
snapshot_download("yogya2005/rs-imitation-learning-data", repo_type="dataset",
                  local_dir="data", allow_patterns=["selfplay/*"])

How it was made

• Simulator: grSim, Division B field (9 × 6 m), 6 robots per team.
• Expert: a fork of Sumatra whose exporter writes these logs (rl-framework/ in the repo).
• Self-play: scripts/collect_selfplay.sh. Play is restarted automatically when the ball stops or leaves the field, or when the game stalls.
• DAgger: scripts/dagger_night.sh, 4 matches of 5 minutes per round.

Licence and attribution

The expert behaviour in these logs comes from Sumatra by TIGERs Mannheim. Its licence permits educational, research and non-commercial use, including RoboCup participation.