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

Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models

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 VLAct, a representation-centric continued pre-training recipe for VLM-based VLA backbones (shallow-layer freezing, caption co-training, multi-head continuous action supervision, partially unified action layout) that reaches competitive results on LIBERO-Plus, RoboTwin 2.0, RoboDojo and RoboCasa-GR1 using 16 GPUs and open data.

Problem

Under a fixed robot-data budget, VLA continued pre-training tends to over-specialize the VLM backbone to a single action head and embodiment and erodes general vision-language features; the goal is a reusable, transferable VLA backbone.

Approach

Continued pre-training of a Qwen3-VL-4B backbone with frozen vision encoder and lower LLM half, caption co-training, simultaneous OFT/PI/GR00T action heads on a shared latent, and a partially unified cross-embodiment action space with wrap-aware angle loss; the pre-training heads are discarded at fine-tuning.

Key results

• VLAct reaches 82.6% on LIBERO-Plus versus 75.0% for the in-house Qwen3VL-OFT baseline and 80.5% for Abot-M0.
• On RoboTwin 2.0 Data Scaling setting, VLAct-OFT reaches 92.5% Clean and 90.8% Random success.
• On RoboCasa-GR1 with 20% of downstream trajectories, VLAct reaches 49.5%, above full-data GR00T-N1.6 at 47.6%.

Surprising findings

• Pure language instruction data as auxiliary co-training also improves over robot-only pre-training on LIBERO-Plus.
• OFT-only pre-training lowers downstream PI fine-tuning performance below from-scratch (55.1 vs 60.5), indicating head-specific representation collapse.

Limitations

• Memory tasks on RoboDojo remain a clear weakness (0.56% success).
• Comparisons with industrial systems use published numbers under differing training compute and protocols.
• Simulation and real-world experiments are limited in scale; real-world uses 10 trials per task.

Experimental setup

Datasets: DROID InternA1 RoboCoin MolmoAct LLaVA-ReCap-CC3M LLaVA OneVision RefCOCO COCO-ReM PixMo-Points RoboPoint SenseNova-SI-800K Nemotron-SFT-Instruction-Following-Chat-v2

Benchmarks: LIBERO-Plus VLA-Arena RoboTwin 2.0 DOMINO RoboCasa-GR1 RoboDojo

Metrics: success rate average score partial-progress score

Baselines: Qwen3VL-OFT pi0 pi0.5 Abot-M0 GR00T-N1.6 LingBot-VLA X-VLA OpenVLA-OFT StarVLA-alpha X-WAM

Builds on: StarVLA OpenVLA-OFT pi0 GR00T N1.5 FAST

Closest papers by meaning

• PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
• FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.92)
• Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better (NeurIPS 2025 · similarity 0.92)
• NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
• VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)