Venue: ICLR 2026 Authors: Shanghai AI Lab + USTC + SJTU + Zhejiang + Nanjing + Fudan + Tsinghua + NUS + Northeastern + Shenzhen U. (Ganlin Yang, Tianyi Zhang, Haoran Hao, Weiyun Wang, Yibin Liu, … Yao Mu, Zhi Hou) Category: VLA Architecture — Reasoning-augmented · VLM-init study Trend tag: Embodied reasoning · VLM→VLA initialization · domain-shift mitigation

flowchart LR
  Net[InternVL3 backbone<br/>2B / 8B] --> SFT[Stage 1: VLM SFT<br/>Vlaser-6M dataset]
  D6M[Vlaser-6M:<br/>1.5M grounding · 1.2M RoboVQA<br/>0.5M spatial · 0.4M planning<br/>2M in-domain SimplerEnv + RoboTwin2.0] --> SFT
  SFT --> VLM[Vlaser VLM<br/>SOTA on 12 embodied benchmarks]
  VLM --> AE[Action Expert<br/>flow-matching, π0-style<br/>shared self-attention<br/>non-causal VLA stream]
  Obs[image + lang + robot state q_t] --> AE
  AE --> A[Action chunk A_t<br/>H=4, 10 inference steps Euler]

VLAs are typically built by fine-tuning generic web-pretrained VLMs (Kim '24 OpenVLA, Black '24 π₀, Qu '25 SpatialVLA), but few studies examine how the VLM initialization choice affects downstream VLA. The hidden assumption is that reasoning-stronger VLMs make stronger VLAs — Vlaser tests it. They find a counter-intuitive result: out-of-domain embodied reasoning data does NOT translate cleanly to closed-loop manipulation; what helps is in-domain data from the same simulator/embodiment.

• VLM backbone: InternVL3-2B (InternViT + Qwen2.5-1.5B) and InternVL3-8B (InternViT + Qwen2.5-7B). Two model sizes for compute-vs-quality trade-offs.
• Action expert: π₀-style mixture-of-experts head — original VLM weights handle image/text streams; a separate weight set handles robot state + noised actions. Self-attention is shared between language and action experts; non-causal attention for the VLA stream.
• Robot state q_t encoded as a state token; noised actions A_t^τ as action tokens; both projected via linear layer into the same embedding space as image/text features.

Six categories, 6M total samples:

The "in-domain 2M" subset is the key empirical lever — synthesized QA annotations directly on robot observation footage from the SimplerEnv simulator (WidowX + Google Robot) and RoboTwin2.0 (Aloha-AgileX), in the reasoning categories above.

Stage 1 — Vision-Language Pretraining (SFT on Vlaser-6M):

• Auto-regressive language-modeling loss L_lm = -log p(t_N | F_v(x; θ_v), F_t(y), t_{0:N-1}; Θ).
• Initialized from InternVL3-2B / InternVL3-8B; full-parameter SFT.

Stage 2 — VLA fine-tuning (flow matching):

• Action chunk A_t = [a_t, …, a_{t+H-1}] with horizon H = 4 (Action Chunk length = 4, Table 7).
• Noisy action chunk A_t^τ = τA_t + (1-τ)ε, ε ~ N(0, I).
• Loss L_vla = E_{p(A_t|o_t)} ‖v_θ(A_t^τ, o_t) - u(A_t^τ | A_t)‖² where u = ε - A_t.
• Inference: forward Euler numerical integration, 10 inference steps; observation history length = 1 (single-frame). At eval only 2 of the 4 predicted actions are executed open-loop before re-planning (Execute Action length = 2, Table 7 / eval appendix) — note the Table 5 ablation separately reports a default "execute length H=4", an internal terminology clash in the paper.
• Eval on SimplerEnv (WidowX Bridge + Google Robot) and RoboTwin 2.0 (Aloha-AgileX bimanual, Table 4). No real-robot experiments.

Training hyperparameters (Appendix Tables 6–7): VLM SFT — AdamW, peak LR 2e-5, cosine decay, 1 epoch / 5,000 steps, global batch 128, 150 warm-up steps, bf16, LLM seq len 16,384, dynamic resolution (patch 448, max 12 patches), all params trainable. VLA fine-tuning — AdamW, VLM & action-expert peak LR 5e-5, global batch 1,024, 10 epochs, seq len 384, all params trainable; WidowX checkpoint at 45,390 iters, Google Robot at 36,970 iters. (GPU hours not reported.)

Vlaser-2B avg 45.3 beats GPT-4o (34.2) and Claude-3.7-Sonnet (33.6); Vlaser-8B avg 51.3 beats Gemini-2.5-Pro (44.4). +30 pt jump from InternVL3-2B init (15.2 → 45.3) and +29 pt from InternVL3-8B (22.3 → 51.3) demonstrates Vlaser-6M's effectiveness.

SOTA on WidowX (Vlaser-All 65.1 vs π₀'s 54.9 with similar param count). Crucially, the Vlaser-OOD (2B) row trained on out-of-domain Vlaser-6M reasoning data alone is only marginally better than InternVL3-2B (43.2 vs 41.8) — adding in-domain Bridge QA/Spatial/Grounding subsets does the heavy lifting (each ≈ +17-21 pts; Vlaser-QA 62.6, combining all three reaches 65.1).

Best Vlaser variant (Vlaser-All, all three in-domain streams combined) reaches VM avg 76.2 and VA avg 59.0 — beating π₀ on VM by +17.9 pts (76.2 vs 58.3). Magma (8B) still leads VA (62.6) via a much stronger Drawer task; Vlaser leads on Pick/Move-Near.

Same pattern: Vlaser-OOD ≈ InternVL3 baseline; in-domain data drives gains. Best per-task: Click bell 92%, Shake bottle 96%, Place mouse pad 92%.

Default (P=4, H=4, 10 steps) is best; doubling sample steps gives marginal gain. Lower H breaks the InternVL baseline harder than it breaks Vlaser — suggesting Vlaser's enhanced state perception is more robust to short execution chunks.

The clean ablation split Vlaser-6M into OOD reasoning only (Vlaser-OOD), +QA, +Spatial, +Grounding, and All (the three in-domain types combined). All three in-domain types help individually (≈ +17-21 pts on WidowX vs the 43.2 OOD baseline: QA 62.6, Spatial 60.8, Grounding 62.0), and combining all three pushes WidowX to 65.1% and RoboTwin to 67.5% — diminishing-returns but additive.

1. No real-robot experiments (the single limitation explicitly stated by the authors). Only simulation (SimplerEnv, RoboTwin) so far; real-world validation deferred to future work.
2. (Reviewer-inferred) Domain gap analysis is descriptive, not prescriptive. Authors observe internet-scale embodied reasoning data ≠ closed-loop control success but don't yet propose a complete fix; in the discussion they argue future work needs alignment techniques to bridge robot-viewpoint vs internet-data viewpoint.
3. (Reviewer-inferred) Narrow hyperparameter exploration — only 4 (P,H,δ) combinations × 3 model variants tested.

Vs OpenVLA / π₀ / SpatialVLA / RoboVLM (action policy peers). OpenVLA effectively scores 1% on WidowX SimplerEnv (very poor). Vlaser-All-2B at 65.1% beats π₀-3B (54.9%) and SpatialVLA-4B (42.7%) at smaller scale. On Google Robot VM, Vlaser-All beats π₀ by +17.9 pts. Cleanest claim: with the right embodied-reasoning init, a 2B model outperforms 3-8B competitors.

Vs MolmoAct / Embodied-R1 / RoboBrain2.0 / Cosmos-Reason1 (embodied-VLM peers). These work on the upstream VLM but most don't ship a downstream VLA. Vlaser is the first to systematically study upstream→downstream transfer and ship both. Beats Embodied-R1-7B by +12.4 pt avg reasoning, beats RoboBrain2.0-7B by +14.3 pt avg.

Vs GR00T-N1.5 (humanoid foundation). Similar size (Vlaser-2B vs GR00T 2.1B), similar pretraining philosophy, but Vlaser focuses on bimanual + single-arm manipulation in 2 simulators while GR00T targets humanoid full-body. Vlaser beats GR00T on WidowX-style benchmarks; GR00T's strength is humanoid-specific data scale.

Empirical contribution to community knowledge. The "in-domain QA helps, out-of-domain reasoning doesn't" finding contradicts the prevailing intuition that better VLM benchmark scores → better VLA. This shifts the design lever for future VLAs from "scale embodied-reasoning data" to "align VLM observation distribution with the target robot's viewpoint". The argument is that the visual-domain mismatch (internet third-person → robot front-view simulation) bottlenecks transfer more than reasoning capability does.

Vs Embodied-R1 / OneTwoVLA (reasoning-augmented VLAs). Those add reasoning into the action policy itself. Vlaser argues for front-loading reasoning into the VLM stage then doing standard flow-matching action expertise — keeps inference fast and policy clean.

Open release of weights, data engine, and the full Vlaser-6M dataset is non-trivial — the dataset itself becomes a community resource on par with Vlaser-6M's competitor curation efforts (Cosmos-Reason1, RoboBrain2.0, EmbodiedOneVision).

• OpenReview: https://openreview.net/forum?id=8xTDnj39Ti
• GitHub: https://github.com/OpenGVLab/Vlaser/

• OneTwoVLA
• InstructVLA
• Embodied R1
• Spatial Forcing

← Back to ICLR-2026