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

ChainFlow-VLA: Causal Flow Planning with Vision-Language Models

NeurIPS 2026 · 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 ChainFlow-VLA, which generates K causal AR trajectory modes and refines them with a VLM-conditioned residual DiT diffusion module, reaching PDMS 94.85 on NAVSIM v1.

Problem

End-to-end driving planners either generate trajectories autoregressively (causal but error-accumulating) or with diffusion (global but lacking explicit causal structure), and existing VLM integrations inject semantics too early to help fine-grained trajectory refinement.

Approach

Factorizes planning as a mixture over AR-induced trajectory modes, with a diffusion refiner that models residuals around each AR proposal and is conditioned on frozen VLM hidden states (cross-attention plus adaLN), so the VLM modulates correction rather than generation.

Key results

• ChainFlow-VLA (trainval) reaches PDMS 94.8 on NAVSIM v1, versus 93.8 for the best prior end-to-end method (RAP-DINO, which uses a larger private pretraining set) and 92.4 for the best VLA baseline (LatentVLA).
• Ablation on DrivoR backbone: AR generator raises PDMS 93.7 to 94.0, adding DiT refiner to 94.1, and VLM guidance to 94.8.
• Increasing denoising steps from 2 to 12 raises PDMS from 94.68 to 94.85; default is 4 steps (94.72).

Surprising findings

• The reported 94.85 at 12 denoising steps matches the human expert PDMS of 94.8, which the authors highlight as human-level performance.
• Residual-space modeling outperforms direct trajectory-space prediction (94.72 vs 92.89 on trainval).

Limitations

• VLM guidance uses a general driving-oriented VLM not optimized under the diffusion objective; a judge-oriented VLM may be better aligned.
• Gains over the DrivoR baseline are modest (about 1.1 PDMS points), and the trainval setting is needed for the headline number.
• Reported human-level score comparison relies on the NAVSIM human-driver PDMS reference.

Experimental setup

Datasets: NAVSIM v1 navtrain NAVSIM v1 trainval

Benchmarks: NAVSIM v1

Metrics: PDMS NC DAC EP TTC Comfort

Baselines: UniAD VADv2 iPad DiffusionDrive Hydra-MDP++ Centaur DriveSuprim RAP-DINO DrivoR UniVLA AutoVLA FSDrive

Builds on: ReCogDrive DiffusionDrive DrivoR DiT NAVSIM

Closest papers by meaning

• OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation (NeurIPS 2026 · similarity 0.90)
• AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.90)
• SOLVE: Synergy of Language-Vision and End-to-End Networks for Autonomous Driving (CVPR 2025 · similarity 0.89)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.89)
• $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.89)
• From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving (NeurIPS 2026 · similarity 0.89)