Skip to content

OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

OneTwoVLA: unified VLA with adaptive reasoning-acting switching for robotics

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

OneTwoVLA is a single unified vision-language-action model that collapses the dual-system (reasoning + acting) paradigm into one model by adaptively switching between explicit language reasoning and direct action generation via learned decision tokens. It outperforms both flat VLA baselines (by 30%) and dual-system approaches (by 24%) on long-horizon manipulation tasks. A scalable synthetic data pipeline using Gemini 2.5 Pro + FLUX further unlocks generalization to novel instructions unseen during robot training.

Problem

Dual-system robot architectures (VLM as System Two + VLA as System One) suffer from three compounding failure modes: (1) lack of mutual awareness between systems — System Two may issue commands System One cannot execute; (2) inference latency causing outdated reasoning guidance; and (3) limited co-training synergy because the two systems are disjoint. Existing unified approaches either reason at every step (costly) or skip reasoning at test time (ECoT-Lite, degraded performance).

Method

OneTwoVLA extends an existing VLA backbone (\(\pi_0\), flow-matching action expert + autoregressive VLM) with two special decision tokens: [BOR] (beginning of reasoning) and [BOA] (beginning of action). At each timestep, the model conditions on current multi-camera observations \(I_t^{1:n}\), reference images from the last reasoning step \(I_\text{ref}^{1:n}\), language instruction \(\ell\), and the latest reasoning content \(R\), then predicts either [BOR] or [BOA].

  • Reasoning mode ([BOR]): auto-regressively generates structured text \(\hat{R}\) covering scene description, high-level plan, historical summary, and next step; supervised with cross-entropy loss.
  • Acting mode ([BOA]): generates action chunk \(A_t\) conditioned additionally on proprioceptive state \(s_t\); supervised with flow-matching loss.

Robot trajectories are segmented into reasoning intervals (annotated with textual reasoning via a two-stage Gemini 2.5 pipeline: interval localization → reasoning content generation) and acting intervals. A scalable VL synthesis pipeline generates 16,000 image–instruction–reasoning samples using Gemini 2.5 Pro for layout descriptions and task generation, and FLUX.1-dev for image synthesis, with fisheye/gripper augmentation to close the sim-to-real gap.

Key Contributions

  • Unified VLA that adaptively decides when to reason vs. act via learned [BOR]/[BOA] tokens, achieving flat-VLA inference speed while retaining reasoning capability.
  • Automated two-stage embodied reasoning annotation pipeline using Gemini 2.5 on demonstration trajectories (81.5% interval accuracy, 83.3% scene description quality by human evaluation).
  • Scalable synthetic VL data pipeline (16K samples) enabling generalization to novel task instructions never seen in robot training data.
  • Demonstrated error detection/recovery, natural human-robot interaction (mid-execution instruction updates, ambiguity clarification), and generalizable visual grounding as emergent capabilities within a single model.

Results

  • Long-horizon planning (Tomato-Egg, Hotpot, Cocktail; 20 trials each): OneTwoVLA avg. 87% success vs. \(\pi_0\) flat VLA 57% (↑30%) and dual-system approach 63% (↑24%).
  • Per-task: Tomato-Egg 95% / Hotpot 80% / Cocktail 85%.
  • Generalizable planning (OneTwoVLA-VL with VL co-training): avg. 65% vs. OneTwoVLA without VL data 6% and \(\pi_0\) 5%, on four novel tasks (Get Icy Cola 70%, Empty Plate 70%, Tool Use 60%, Prepare Drinks 60%).
  • Inference latency: OneTwoVLA total task time matches flat \(\pi_0\) (176 s vs. ~176 s) vs. dual-system 260 s on Tomato-Egg, because reasoning is invoked only at critical steps (~109 s reasoning time overlaps with execution).

Limitations

  • Evaluated exclusively on tabletop manipulation with a single Franka arm (and limited dual-arm experiments); generality to mobile manipulation or non-tabletop settings is unverified.
  • Synthetic VL data generation pipeline relies on proprietary Gemini 2.5 Pro and FLUX.1-dev; reproducibility and cost at scale are not analyzed.
  • Error recovery and human-robot interaction are demonstrated qualitatively but lack quantitative success-rate ablations.
  • Reasoning interval annotation requires a predefined high-level plan \(P = (p_1, \ldots, p_K)\) per task, introducing manual structure that may not generalize to open-ended tasks.
  • The 16K VL samples cover tabletop household objects; domain shift to industrial or outdoor settings is unaddressed.

Relevance to Vision-Language Models

OneTwoVLA demonstrates that the generalization and instruction-following capabilities of VLMs transfer directly to embodied control when reasoning and acting are unified within a single model rather than decoupled. The adaptive [BOR]/[BOA] decision mechanism is a concrete instantiation of how a VLM can self-regulate its own "thinking budget" — a question of broad interest in efficient inference for VLMs. The synthetic co-training pipeline (VLM-generated layout descriptions → text-to-image synthesis → VLM-annotated tasks) is a scalable recipe for grounding pretrained VLM knowledge into action-space domains, which generalizes beyond robotics to any task requiring VLM-to-actuator grounding. The work reinforces the trend that co-training on diverse VL data (not just task-specific demonstrations) is critical for out-of-distribution generalization in VLA models.