MJ22x/go-vla-benchmark

Go VLA Benchmark — Model Checkpoints & Evaluation Results

Fine-tuned VLA (Vision-Language-Action) models on a simulated 5×5 Go stone placement task using a Panda robot arm in robosuite/MuJoCo.

Task

Each episode: pick up a stone from a source tray and place it at a specific intersection on a 5×5 Go board. The instruction specifies the stone color and target position (e.g., "Place a black stone on the Go board at row 3, column 1.").

Models

Model	Base	Params	Fine-tuning	Checkpoint Size
OpenVLA	openvla/openvla-7b	7B	QLoRA (r=32, 4-bit)	~460MB adapter
SpatialVLA	IPEC-COMMUNITY/spatialvla-4b-224-pt	4B	LoRA (r=32)	~260MB adapter
π₀ (pi-zero)	pi0_base (PaliGemma 3B + action expert)	3.3B	Full fine-tune	~12GB per checkpoint

Training Data

100 scripted demonstrations collected in robosuite with domain randomization (board position, lighting, camera, textures, stone color). Converted to RLDS format (v5.0.0). 98 train / 2 val split.

Evaluation Protocol

For each model (best + latest checkpoint):

1. 8 train-seed rollouts — deterministic seeds matching training-like board configs
2. 2 OOF rollouts — out-of-fold with high opening moves
3. 10 eval rollouts — random board configurations
4. 5-move self-play — random move selection, force-commit on failure

All rollouts run in the actual physics engine with compounding error. Max 200 steps per episode.

Results

Model	Checkpoint	Train SR	OOF SR	Eval SR	Self-play
OpenVLA	best	0%	0%	0%	0/5
OpenVLA	latest	0%	0%	0%	0/5
SpatialVLA	best	0%	0%	0%	0/5
SpatialVLA	latest	0%	0%	0%	0/5
π₀	best	0%	0%	0%	0/5
π₀	latest	0%	0%	0%	0/5

Note: All models achieve low L1 action error on isolated frames (no compounding error) — OpenVLA: 0.039, π₀: 0.083 — but fail to complete the task in the physics engine due to compounding error over the ~70-step trajectories.

Directory Structure

openvla/
  best/          — best checkpoint by val loss (LoRA adapter)
  latest/        — final checkpoint (LoRA adapter)
  wandb-summary.json

spatialvla/
  best/          — best checkpoint (LoRA adapter)
  latest/        — final checkpoint (LoRA adapter)

pizero/
  best/          — best checkpoint (full model weights)
  latest/        — final checkpoint (full model weights)
  wandb-summary.json

eval_summary.json  — aggregated results for all models

Each checkpoint contains:
  eval/
    results.json      — metrics
    train_00..07.mp4  — train-seed rollout videos
    oof_00..01.mp4    — out-of-fold rollout videos
    eval_00..09.mp4   — random eval rollout videos
    self_play.mp4     — 5-move self-play video

Training Details

• OpenVLA: 20 epochs, QLoRA r=32, 4-bit quantization, batch 128 (4×32 grad accum), lr=5e-4, 500 gradient steps
• SpatialVLA: 50 epochs, LoRA r=32, batch 256 (32×8 grad accum), lr=5e-4, DeepSpeed ZeRO-1
• π₀: 2900 steps, full fine-tune, batch 32, lr=2.5e-5, cosine decay

Environment

• robosuite 1.5.2 with rigid-body Go stones
• Panda arm, OSC position controller
• 256×256 agentview camera
• Domain randomization: board position (±1.5cm), lighting, camera FOV/angle, table/board/stone textures and colors