Research map / Vision-language and multimodal / Robot action and navigation

Fast-in-Slow: A Dual-System VLA Model Unifying Fast Manipulation within Slow Reasoning

NeurIPS 2025 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Fast-in-Slow (FiS-VLA), a dual-system VLA that repurposes the final transformer blocks of a 7B VLM (LLaMA2 backbone) as a high-frequency System 1 diffusion action head inside the intact System 2 VLM, trained with a dual-aware co-training loss.

Problem

Vision-language-action models built on billion-parameter VLMs run at low control frequency, while existing dual-system designs use separate lightweight System 1 heads that do not fully exploit the VLM's pretrained knowledge.

Approach

Embed the System 1 execution module as the final few transformer blocks of the same VLM used for System 2 reasoning (partial parameter sharing), run System 2 at a lower frequency (1:4 ratio) and System 1 at high frequency with heterogeneous inputs (state, 2D images, 3D point-cloud tokens), and co-train with a diffusion denoising loss for System 1 plus an autoregressive loss for System 2.

Key results

• On RLBench, FiS-VLA reaches a mean success rate of 0.69 versus 0.61 for CogACT and 0.55 for pi0, with 21.9 Hz inference at action chunk 1.
• With action chunk 8 and a 1:4 frequency ratio, FiS-VLA reaches a theoretical 117.7 Hz on an NVIDIA 4090.
• Real-world mean success is 0.68 vs 0.59 for pi0 on Agilex and 0.74 vs 0.61 on AlphaBot.

Measured effects reported in the paper

Surprising findings

• Removing the autoregressive slow loss L_slow drops RLBench success from 69% to 62%, suggesting co-training preserves useful System 2 reasoning for System 1.
• Performance saturates at two shared blocks; eight shared blocks (0.64) do not beat two (0.69).

Limitations

• System 1 shared parameters and the System 2/System 1 frequency ratio are statically configured rather than adaptive.
• Failure cases include bimanual collisions, manipulation height errors, and mispredicted object positions.
• Real-world evaluation is limited to eight tasks on two dual-arm robots with 20 rollouts per task.

Experimental setup

Datasets: Open X-Embodiment DROID RoboMIND RLBench RDT pretraining mixture

Benchmarks: RLBench (10 tasks) real-world Agilex dual-arm tasks real-world AlphaBot dual-arm tasks

Metrics: success rate control frequency (Hz)

Baselines: ManipLLM OpenVLA pi0 CogACT

Builds on: CogACT Prismatic VLM OpenVLA

Closest papers by meaning

• VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.90)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.89)
• ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.89)
• RoboCerebra: A Large-scale Benchmark for Long-horizon Robotic Manipulation Evaluation (NeurIPS 2025 · similarity 0.89)
• VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)