Recoleta
GitHub
Source note

CrossVLA: Cross-Paradigm Post-Training and Inference Optimization for Vision-Language-Action Models

Authors
Zhi Liu
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionRobot Policy AlignmentFlow MatchingDpoParameter Efficient Finetuning
Open arXivMarkdown source

Summary

CrossVLA studies DPO post-training across autoregressive and flow-matching vision-language-action models, with a practical surrogate log-probability for continuous action chunks. Its strongest claims are DoRA+DPO gains on OpenVLA LIBERO tasks and a negative result showing prefix KV caching is a poor speedup target for flow-matching VLAs.

Problem

  • VLA post-training has focused on autoregressive action-token models such as OpenVLA, while flow-matching models such as pi_0.5 lack a cheap log-probability needed for DPO.
  • Robot policy adaptation needs methods that work across discrete-token and continuous-action backbones without full retraining.
  • Inference speed work for VLAs often targets prefix KV caching, but flow-matching policies spend most time in action denoising.

Approach

  • For flow-matching VLAs, the paper replaces exact chunk log-likelihood with negative flow-matching MSE over sampled time points, using it as a surrogate log-probability inside standard DPO.
  • It defines a common VLA interface for log-probability, reference log-probability, sampling, observation encoding, and action generation across OpenVLA and pi_0.5.
  • It compares LoRA and DoRA as parameter-efficient DPO adapters, with DoRA separating weight magnitude and direction.
  • It measures pi_0.5 inference latency by stage and tests chunk-level caching plus token-level prefix KV caching.
  • It also trains a frozen-SigLIP projection head with multi-view and temporal InfoNCE on LIBERO frames.

Results

  • On OpenVLA over LIBERO 4-suite, DoRA+DPO reaches 73.2% mean success versus 62.75% SFT, a +10.4 percentage point gain across 600 trials and 3 environment seeds.
  • Per-suite DoRA+DPO gains over OpenVLA SFT are Object +20.0 pp, Long-horizon +11.0 pp, Goal +8.0 pp, and Spatial +2.7 pp; Object is 76.0% across all three seeds, 38/50 each time.
  • Compared with LoRA where multiseed results are available, DoRA is Object 76% vs 75%, Goal 78% vs 77%, and Long-horizon 64% vs 64%.
  • The flow-matching DPO surrogate trains stably on pi_0.5; completed LIBERO Spatial and Object runs stay at their saturated SFT levels, 100% and 98%.
  • pi_0.5 latency is about 280 ms per sample_actions call: image preprocessing about 5 ms, prefix forward about 60 ms or 21.4%, and the 10-step denoise loop about 220 ms or 78.6%.
  • Prefix-style caching performs poorly on pi_0.5: chunk caching drops LIBERO Spatial success from 50/50 to 40/50 and increases wall time from 1258 s to 1796 s; token-level prefix caching reports failed runs at 0/1 and 0/2 under cache-hit settings. The projection-head pretraining reaches 99.5% k-NN recall@1 on 6000 LIBERO frames, 36x random.

Link