Skip to content

Latest commit

 

History

History
54 lines (40 loc) · 2.39 KB

File metadata and controls

54 lines (40 loc) · 2.39 KB

OpenVLA — Reproduction Report

Open-source VLA model. GitHub | Paper | 7B params (Prismatic VLM + action head).

Results Summary

Benchmark Reproduced Reported Verdict
LIBERO 76.2% 76.5% Reproduced

LIBERO

Checkpoints Per-suite fine-tuned (LoRA), see below
Server configs configs/model_servers/openvla/libero_*.yaml
Benchmark configs configs/benchmarks/libero/spatial.yaml, etc.
Results data/openvla-libero/

4 suites × 10 tasks × 50 episodes = 2000 episodes. Per-suite fine-tuned checkpoints (LoRA).

Suite Checkpoint Reproduced Reported
Spatial openvla/openvla-7b-finetuned-libero-spatial 86.6% 84.7%
Object openvla/openvla-7b-finetuned-libero-object 86.6% 88.4%
Goal openvla/openvla-7b-finetuned-libero-goal 79.0% 79.2%
Long openvla/openvla-7b-finetuned-libero-10 52.8% 53.7%
Average 76.2% 76.5%

Per-suite reported scores from Table I of the paper (arxiv 2406.09246).

Pipeline Notes

Matching the reference eval required aligning image preprocessing and env setup. The reference (openvla/experiments/robot/) applies JPEG roundtrip, Lanczos3 resize to 224×224, center crop (scale=0.9), and env.seed(0).

  • Center crop — isolated impact ~3pp (73.3% → 76.2%).
  • JPEG roundtrip, env_seed=0, Lanczos resize — applied together; individual contributions not isolated. Combined with center crop: 71.0% → 76.2%.

Without any of these fixes, the average was 71.0%.

Configuration Notes

  • chunk_size=1 (no action chunking).
  • Per-suite unnorm_key required: libero_spatial, libero_object, libero_goal, libero_10.
  • Gripper: RLDS format [0=close, 1=open] converted to robosuite [-1=open, +1=close] via -(2x - 1).
  • Action output: 7D [pos3, axisangle3, gripper].
  • Each suite requires a different model checkpoint — cannot evaluate all suites with a single model.
  • 7B model requires ~15 GB GPU memory (bfloat16).
  • transformers==4.40.1 required (trust_remote_code model).
  • Single image input (no wrist camera, no proprioceptive state).