4-Bit Robot Policy Reaches Full-Precision Mean Score With QCRL

A robot policy using four-bit weights and activations reached 100.7% of its full-precision reference score, with deployment running 1.73 times faster on Jetson Thor. The result came from a study spanning five vision-language-action models and 4,050 real-world trials, according to an arXiv preprint.

Quantization reduces the numerical precision used by vision-language-action, or VLA, policies and can lower deployment costs. The study reports that it can also compromise task completion.

The researchers found that quantization-aware training generally performed better than post-training quantization. It still left substantial gaps against full-precision policies on long-horizon tasks and under visual and semantic shifts.

The proposed method, called Quantization-Conditioned Reinforcement Learning, or QCRL, is an offline reinforcement-learning approach. Unlike the imitation-based objective discussed in the paper, QCRL explicitly uses rollout success and partial-completion signals. Full-precision rollouts provide reference experience, while a progress-scaled terminal penalty distinguishes levels of partial completion. Quantization-aware conditioning selects advantageous reference behavior within the quantized policy.

On the π0.5 model, an INT4 configuration with four-bit weights and 16-bit activations reached 99.3% of the full-precision model’s mean score across three visual-shift conditions. The result used only in-distribution data. The same configuration reduced model memory by 42.9% relative to FP16 on Jetson Thor.

The researchers also evaluated weight-only INT4 compression on GR00T N1.5 and a joint four-bit weight-and-activation format, NVFP4 W4A4, on π0.5. The NVFP4 version reached 100.7% of the same full-precision baseline and delivered the reported 1.73-times speedup using NVFP4/FP8 deployment. The hardware gains were measured relative to FP16 on Jetson Thor.

Those percentages describe evaluation scores and do not establish that every compressed-policy attempt succeeded. The supplied report does not disclose the individual manipulation tasks, absolute completion rates, latency figures, power draw or how the 4,050 trials were distributed among the five models and compression settings. The results indicate that quantization can impair VLA policy performance and that QCRL recovered much of the reported loss by incorporating rollout outcomes and partial completion.