Kris Wu

VLA Dev Team Lead @airoa_org | prev Research Scientist @nvidia | Ph.D. @UCSanDiego

[Kris Wu @yh_kris]

VLA generalization is hard in part because we ask a robot policy to solve two under-specified problems at once, with far less data than today’s VLMs.

From coarse goals and full images alone, it must infer both what to do now and what in the scene matters for control, then turn that into precise actions from scarce, embodiment-specific robot data.

In See Less, Specify More, we reduce that burden. S2 uses an off-the-shelf VLM for broader reasoning and goal-preserving local guidance, including trajectory-aware execution details, then trains the action policy to act from annotation-free, task-sufficient visual evidence instead of broad scene context.

Across 8 real-robot tasks on TX-G2 and HSR, S2 raises mean subtask success from 54.2% to 79.0% over pi05

[Kris Wu @yh_kris]

Not every latent is reasoning.

We’re releasing Continuous Reasoning for Vision-Language-Action.

Our claim is simple: if a robot policy is truly reasoning, another policy should be able to decode the same thought, verify it, and use it to make better action predictions.

(1/6)

[Kris Wu @yh_kris]

@stevencheng Thanks! That data wall is exactly what motivated this work.

[Kris Wu @yh_kris]

VLA generalization is hard in part because we ask a robot policy to solve two under-specified problems at once, with far less data than today’s VLMs.

From coarse goals and full images alone, it must infer both what to do now and what in the scene matters for control, then turn that into precise actions from scarce, embodiment-specific robot data.

In See Less, Specify More, we reduce that burden. S2 uses an off-the-shelf VLM for broader reasoning and goal-preserving local guidance, including trajectory-aware execution details, then trains the action policy to act from annotation-free, task-sufficient visual evidence instead of broad scene context.

Across 8 real-robot tasks on TX-G2 and HSR, S2 raises mean subtask success from 54.2% to 79.0% over pi05

[Kris Wu @yh_kris]

Exactly! That is very close to our view. The policy is often asked to infer both the “what” and the “how” from full visual context, but with far less data than modern VLMs. S2 tries to reduce that burden from both sides: more specific local guidance on the “what,” and annotation-free evidence budgeting on the “how.��

[Kris Wu @yh_kris]

Thanks! That is exactly the bottleneck we were trying to target.

On noisy local cues, HSR already gives us a fairly challenging case: during grasping, the wrist camera is often partially occluded by the grasped object itself, which reduces the available light and makes the wrist-view signal quite noisy. Even so, we still learn a strong policy there, especially on tasks that require two locomotion phases plus relocation.

On hardware changes, so far we have trained separate multi-task policies on HSR and TX-G2, and the gains remain strong within each embodiment. A unified policy across embodiments — and understanding how different embodiments learn different visual evidence under the same environment — is definitely a very interesting next direction for us.

[Kris Wu @yh_kris]

With @__tmats__ and @ohtake_i

arXiv: https://t.co/pnj1E3Fs2p
Project page: https://t.co/srXUAwOYhu

[Kris Wu @yh_kris]

This is also my first work at @airoa_org , with @__tmats__ and @ohtake_i .

arXiv: https://t.co/Yl90GRO4lO
Project page: https://t.co/NG4chRJSWB

[Kris Wu @yh_kris]

Not every latent is reasoning.

We’re releasing Continuous Reasoning for Vision-Language-Action.

Our claim is simple: if a robot policy is truly reasoning, another policy should be able to decode the same thought, verify it, and use it to make better action predictions.

(1/6)

[Kris Wu @yh_kris]

Continuous Reasoning is not only effective on TX-G2, a bimanual manipulation platform. It also works on HSR, where the policy must combine locomotion, navigation, transport, and placement.

Across real-robot evaluation, it improves mean subtask success over pi05 by:

• +40.4% on TX-G2
• +26.3% on HSR

[Kei Ota @ohtake_i]

AIRoAに移ってから初のICRA参加します！今年は
- メイン会議2本: Touch2Insert, Simultaneous Extrinsic Contact and In-hand Pose Estimation
- WS1本: YUBI (UMI-variant)
- WS主催: VLA Pipelines for Real Robots
- AIRoA Booth: YUBIのデータ収集体験と活動紹介動画
- AIRoA主催Networking dinner

[Kei Ota @ohtake_i]

昨年1年間HSRを用いて進めてきたデータ収集プロジェクトの動画も公開しました。
ゼロから実世界でデータ収集・学習・デプロイ・評価のPDCAを回し続けるのは非常に大変でしたが、多くの知見が蓄積され現在の双腕ロボットでのデータ収集にも活きています。
近日中に5000時間規模の��ータも公開します。🔥

[一般社団法人AIロボット協会 @airoa_jp]

【動画公開】
AIRoAでは双腕モバイルマニピュレータをはじめとする多種多様なロボットやデバイスによるデータ収集、ロボット基盤モデルの開発を行っています。
並行して、汎用セミヒューマノイドの開発コンペも実施しています。

https://t.co/ir5DFKUoSl

※ 本事業はNEDO(国立研究開発法人 新エネルギー・産業技術総合開発機構)からの採択を受けて実施しています。

[一般社団法人AIロボット協会 @airoa_jp]

【動画公開】
AIRoAは、家庭でのロボット活用に向けた実証実験を進めています。今回、開発拠点と家庭環境で動くロボットの様子をまとめた動画を初公開します。
https://t.co/tNpYa2FwwK
#ICRA2026

[Kris Wu @yh_kris]

🏁 Call for Competition Participants
Our competition focuses on evaluating performance in realistic robotic settings. 🦾

Please check our website for more details!
https://t.co/R9H3NttgLE

[Kris Wu @yh_kris]

Excited to announce the ICRA 2026 Workshop on Visual-Language-Action (VLA) Pipelines in Vienna!

Building effective VLA systems involves complex, multi-stage pipelines. Join us to discuss the future of robotic intelligence! 🤖 (1/n) #ICRA2026

[Kris Wu @yh_kris]

📣 Call for Papers
We welcome contributions on any component of the VLA pipeline:
📊 Data & Representations
🧠 Models & Training
✅ Evaluation & System Integration
🔗 End-to-end connections

Submit your novel methods, benchmarks, and applications!