Sitelet https://github.com/Rick0525
Skip to content
View Rick0525's full-sized avatar
1 follower · 3 following

Block or report Rick0525

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Rick0525/README.md

Hi, I'm Rick 👋

I'm a master's student working on embodied AI / Vision-Language-Action (VLA) models: rigorous evaluation, per-episode failure attribution, and RL post-training. My earlier background is in LLM engineering.

This summer I built two projects on a shared 3×A800 server:

  • vla-libero-eval: evaluates open-source VLA models (SmolVLA, π0.5, OpenVLA-OFT) on LIBERO-Spatial under one fixed protocol, with a same-scale leaderboard, per-episode failure attribution (perception / planning / control, every failure episode reviewed on video), and the MuJoCo version-sensitivity finding below. Headline: 87.4 / 94.4 / 97.4 (SmolVLA 0.45B / π0.5 3.3B / OpenVLA-OFT 7.5B) under one protocol, MuJoCo 3.2.7.
  • vla-rl-post-training: reproduces the SimpleVLA-RL GRPO recipe (originally 8×A800 full-parameter) on 2×A800 + LoRA via RLinf: 56.0 → 93.8 on LIBERO-Spatial using ~23% of the official data budget; the official checkpoint scores 96.4 on the same evaluator. Includes a preregistered controlled experiment on a dealing-strategy intervention (speedup confirmed, score gain unproven).

Upstream contributions

Findings from those projects, reported back to the stacks I build on:

  • huggingface/lerobot#4314: PR fixing a dataloader regression where delta queries on datasets ≥ 4.4 fall back to full-row decode (including all image columns); cached column views make it 46× faster per sample, cutting a 30k-step SmolVLA training run from ~4h to ~1.9h. Root-cause analysis in #2895, acknowledged by the HF datasets lead.
  • huggingface/lerobot#4390: MuJoCo version drift silently collapses LIBERO-Spatial task 5 on the same checkpoint (98% → 12% success): a box-box collision bugfix in MuJoCo 3.4.0 breaks the benchmark's pre-3.4 initial states. Verified with a policy-free physics probe (gist) and same-physics A/B evals; cross-reported to LIBERO#141 and RLinf#1460.
  • HuggingFaceVLA/libero · discussion #10: full audit backing a community bug report on the LIBERO dataset re-release; 1690/1693 episodes carry wrong meta pointers, with complete per-split stats plus a working parquet-level workaround.

Contact

📫 rick0525@163.com 💼 Open to embodied-AI / robot-learning internships starting Sept 2026 (part-time during the semester).

Pinned Loading

  1. vla-libero-eval vla-libero-eval Public

    Unified evaluation of three VLA models (SmolVLA / pi0.5 / OpenVLA-OFT) on LIBERO-Spatial with per-failure video attribution; includes three reported upstream findings and the probe tools behind them

    Python 2

  2. vla-rl-post-training vla-rl-post-training Public

    Reproducing 8-GPU VLA RL post-training (SimpleVLA-RL GRPO recipe on RLinf) with 2xA800 + LoRA: 93.8 on LIBERO-Spatial at 23% of the official data budget

    Python 1

  3. No-policy MuJoCo settle probe: LIBER... No-policy MuJoCo settle probe: LIBERO libero_spatial task-5 init states stop seating the bowl on the ramekin from mujoco 3.4.0
    1 
    #!/usr/bin/env python
    2 
    """No-policy physics probe: does the LIBERO libero_spatial task-5 target bowl
    3 
    stay seated on the ramekin after set_init_state + settle, under the CURRENT
    4 
    mujoco version?
    5