Reproduce “Recovering Low-Bit Failures in Vision-Language-Action Policies via Quantization-Conditioned Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seongmin Park, Jaewon Lee, Yujun Heo, Yoonsik Kim, Seheon Choi, Hyoung-rock Kim, Jungwook Choi
QCRL trains quantized VLA policies from offline rollouts using subgoal progress to score failed trajectories and conditioning behavior on advantage and data source. For π₀.₅ at INT4 W4A16, it reaches 99.3% of the full-precision mean across three visual-shift conditions, versus QAT’s 77.5%. This is a ratio of mean task scores, not a success rate; sensor-degradation performance remains below full precision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.