SyVLA

SJTU + Shanghai Innovation Institute + Shanghai AI Lab + Noematrix et al. · 2026-06

VLANew in 20263B VLM + 0.69B expertflow matching

An academic VLA that disclosed its pretraining compute (32 H200s, about 1.5 days).

Paper

Category: New in 2026
Type: VLA
Released: 2026-06
Organization: SJTU + Shanghai Innovation Institute + Shanghai AI Lab + Noematrix et al.
Open weights: Unverified
Parameters: 3B VLM + 0.69B expert (total not stated)

Backbone LLM: Qwen2.5-3B · Qwen2.5-VL-3B config
Vision encoder: Qwen2.5-VL ViT
Action head: flow expert 0.69B (15 layers) (flow matching)
Chunk: 50 steps
Control rate: Not disclosed
Architecture (original summary): Qwen2.5-VL-3B + 0.69B flow matching expert (15-layer decoder, hidden 1408); 20 Feature Query Tokens, 50-step chunks; asynchronous VLM·expert inference

Pretraining accelerators: 32 H200 (DeepSpeed ZeRO-2)
Pretraining duration: About 1.5 days
Pretraining steps & batch: Batch 512 (steps Not stated)
GPU/TPU-hours: 1,152 (Converted · H200-hours (32 GPUs × about 1.5 days = 36 hours))
Fine-tuning: 8 H100 (DeepSpeed ZeRO-2), 6–18 hours, batch 32; RL adds about 10% training time
Data type: Real robot (OXE·RH20T subsets, under 5% self-collected) + web vision-language (VQA)
Data scale: Robot data about 450 hours (25 Hz) + multimodal VQA; CoT annotations about 1%
Main datasets: Open X-Embodiment subset, RH20T subset
Tasks & evaluation: Real shirt folding in-domain 0.86·out-of-domain 0.78; LIBERO Spatial 87.7·Object 84.0·Goal 87.3·Long 65.7
Robot: Not stated in notes

Inference device & speed: Asynchronous VLM·expert inference (latency not reported)
Inference memory: Not disclosed
Weights bf16 (calc.): Not disclosed
Where it runs: Not disclosed

Reinforcement learning: Used (VLA RL post-training) · In stage 3, the whole VLA is optimized with a PPO-variant RL on the real robot (expert sample advantage fixed at 1.0, value model about 100M). Training time rises by only about 10%
RL evidence: https://arxiv.org/html/2606.09009
Base control: Fixed · No mention of base (Cobot Magic arm joint-position actions; LIBERO sim)

Key features: An academic VLA that disclosed its pretraining compute (32 H200s, about 1.5 days). Adding RL increased training time by only about 10%.
Notes: ICML 2026. Project page claims code·dataset release; weight release unverified. Not a Spirit AI model. No matching row in verification notes.
Verification: Not verified