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

D-VLA: A High-Concurrency Distributed Asynchronous RL Framework 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 D-VLA, a distributed asynchronous RL system for VLA models that decouples data and weight-control planes and uses a four-thread "Swimlane" pipeline, dual-pool VRAM management, and topology-aware replication to raise ManiSkill RL training throughput.

Problem

Applying RL to billion-parameter VLA models in large distributed setups is bottlenecked by resource contention between GPU-based physics simulation and deep learning training (VRAM, bandwidth, synchronization), limiting throughput.

Approach

Physically separates high-frequency training data exchange from low-frequency weight synchronization (CPU-side broadcast), overlaps sampling, inference, gradient computation and weight distribution in a four-thread asynchronous pipeline, and isolates simulator memory from model memory in a dual-pool scheme.

Key results

• On pi_0.5 with 16 GPUs, D-VLA (1:1) reaches 336.04 throughput vs 232.23 for RLinf-co in the table (text reports 147.0 vs 127.24 steps/s for single-node), and 3:1 reaches 376.00.
• On OpenVLA-OFT, D-VLA (1:1) reports 250.90 throughput vs 170.48 for RL-VLA^3 (1:1) and 87.20 for RLinf-co in the 16-GPU table.
• Scaling pi_0.5 from 384 to 3,072 environments peaks at 379 steps/s at 768 environments, then declines to around 360 steps/s.

Surprising findings

• Throughput declines beyond 768 environments, attributed to GPU memory bandwidth and compute saturation under high-concurrency rendering.
• Reported numbers are internally inconsistent: text cites 147.0 and 237.0 steps/s for pi_0.5 while the 16-GPU table shows 336.04 and 376.00.

Limitations

• Trillion-parameter scalability is claimed in the abstract but not evaluated in the experiments.
• Learning performance is only shown qualitatively via success-rate curves; the paper does not give quantitative success results against baselines.
• Experiments are limited to ManiSkill with two VLA models, while LIBERO is mentioned but not reported in the experiments.

Experimental setup

Benchmarks: LIBERO ManiSkill

Metrics: throughput (steps/s) step time rollout time actor time success rate

Baselines: RLinf-co RLinf-dis RLinf-hyper RL-VLA^3

Builds on: GRPO FSDP RLinf-VLA RL-VLA^3

Closest papers by meaning

• EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.91)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.89)
• VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.89)