AIRoA (AI Robot Association)

(R&D-029) AI Engineer (VLAs)

Posted 9 Days Ago

Be an Early Applicant

Tokyo, JPN

In-Office

Mid level

Artificial Intelligence • Information Technology • Robotics

The Role

Develop and improve Vision-Language models for robotics, build training pipelines, conduct error analyses, and establish infrastructure for model deployment.

Summary Generated by Built In

※日本語版が続きます。

About AIRoA

The AI Robot Association (AIRoA) is launching a groundbreaking initiative: collecting one million hours of humanoid robot operation data with hundreds of robots, and leveraging it to train the world’s most powerful Vision-Language-Action (VLA) models.

What makes AIRoA unique is not only the unprecedented scale of real-world data and humanoid platforms, but also our commitment to making everything open and accessible. We are building a shared “robot data ecosystem” where datasets, trained models, and benchmarks are available to everyone. Researchers around the world will be able to evaluate their models on standardized humanoid robots through our open evaluation platform.

Job Description

Develop Vision-Language models, or equivalent multimodal models, with a view toward applications in the robotics domain
Fine-tune existing models, conduct evaluations, perform error analysis, and improve performance
Build training pipelines using real-world data, design evaluation metrics, and operate iterative improvement cycles
Prepare and preprocess data, and build training environments for image, video, language, and action data
Research the latest trends in technologies and academic studies, select appropriate technologies, and incorporate findings into model improvements
Establish training infrastructure, inference infrastructure, and experimental environments for real-world model deployment
Collaborate with related teams such as software engineers and robotics engineers to define requirements, design validation plans, and drive development

AIRoAについて

AI Robot Association（AIRoA）は、画期的な取り組みを開始します。数百台のヒューマノイドロボットを用いて、ヒューマノイドロボットの操作データを100万時間分収集し、それを活用してVision-Language-Action（VLA）モデルを学習させます。

私たちの独自性は、実世界データとヒューマノイドプラットフォームの前例のない規模だけではなく、あらゆるものをオープンでアクセス可能にするというコミットメントにもあります。AIRoAはデータセット、学習済みモデル、ベンチマークを誰もが利用できる共有の「ロボット・データ・エコシステム」の構築を目指しています。実現が成功すれば、世界中の研究者が、私たちのオープン評価プラットフォームを通じて、標準化されたヒューマノイドロボット上で自らのモデルを評価できるようになることを期待しています。

業務内容

ロボティクス領域への応用を見据えた Vision-Languageモデル、またはそれに準ずるマルチモーダルモデルの開発
既存モデルの fine-tuning、評価、エラー分析、性能改善
実データを用いた学習パイプライン構築、評価指標設計、改善サイクル運用
画像・動画・言語・行動データ等を対象としたデータ整備、前処理、学習環境構築
最新の研究・技術動向の調査、技術選定、およびモデル改善への反映
モデルの実運用を見据えた学習基盤・推論基盤・実験環境の整備
ソフトウェアエンジニア、ロボティクスエンジニア等の関連チームと連携した要件整理、検証設計、開発推進

Requirements

※日本語版が続きます。

Required Qualifications

Experience leading machine learning models from deployment to improvement and operation in a production service environment
Experience implementing, training, and evaluating machine learning models using Python and PyTorch
Hands-on experience fine-tuning Vision-Language models, or equivalent multimodal models
Experience building training pipelines with real-world data, designing evaluations, conducting error analysis, and operating improvement loops
Ability to understand the latest research and technology trends and translate them into model improvements and practical product applications

Preferred Qualifications

Experience developing Vision-Language-Action (VLA) models or multimodal models for robotics
Experience with robot control, ROS / ROS 2, C++, and real-world hardware evaluation
Knowledge of or experience in sensor integration, actuator control, action generation, and low-level control
Familiarity with training and evaluation using simulators, Sim2Real, and domain adaptation
Experience building training and inference infrastructure in cloud environments such as AWS or GCP
Experience with reproducible and operationally robust development practices such as Docker, CI/CD, and MLOps

必須要件