Software Engineer, Robotics

Deccan AI • San Francisco, CA

177d • Onsite

About The Position

Deccan AI is a model training and evaluation startup (Mountain View, CA; delivery center in Hyderabad). Founded by IIT Bombay, IIM Ahmedabad, and ex-Google alumni and backed by Prosus Ventures, we build expert-curated datasets and evaluation infrastructure for frontier AI labs including Google DeepMind and Snowflake. Our Physical AI practice is building the data backbone for embodied AI — annotation, synthetic data generation, and model evaluation that help robots learn manipulation and reasoning at scale.

The Role
VLA (Vision-Language-Action) foundation models need more than research breakthroughs — they need production software that connects data pipelines, simulation, training loops, and evaluation into a reliable system. You'll build that system. This is not a pure ML research role or a data engineering role — it sits at the intersection, where you write the code that makes VLA training work end-to-end. You'll work directly with frontier lab clients and ship systems used to train the next generation of robot foundation models.

Requirements

• MS or PhD in CS, Robotics, or ML (or equivalent industry experience).
• 2+ years shipping production systems — not just prototypes.
• Strong Python and systems skills.
• Comfortable with Linux, Docker, CUDA, and cloud infrastructure (AWS/GCP).
• Working knowledge of robot learning: imitation learning, behavior cloning, diffusion policies, or RL. You need to understand how VLA training works, not just the theory.
• Experience with at least one robotics sim (Isaac Sim, MuJoCo, PyBullet, Gazebo) and robotics data formats/middleware (ROS/ROS2, URDF/USD, MCAP, HDF5).
• Clean, tested, documented code.
• CI/CD and code review experience expected.

Nice To Haves

• Hands-on with VLA codebases: RT-2, Octo, OpenVLA, π0, GR00T N1, or LeRobot.
• Experience with NVIDIA Omniverse, Replicator, or Cosmos for synthetic data; sim-to-real transfer techniques (domain randomization, NeRF, Gaussian Splatting).
• Built ML eval frameworks, model benchmarking suites, or RLHF/DPO training pipelines.
• Open-source contributions in robotics or ML.
• Published research in manipulation, embodied AI, or VLA systems is a plus.

Responsibilities

• Build the VLA training integration layer — data loaders, format converters, and preprocessing that feed curated datasets (real + synthetic) into training frameworks (LeRobot, Octo, OpenVLA, π0).
• Develop evaluation and benchmarking infrastructure: sim-based rollouts in Isaac Lab, success-rate tracking, regression detection, and automated reporting for client delivery.
• Own dataset management — versioning, schema validation, metadata indexing, and format conversion across Open X-Embodiment, LeRobot HDF5, RLDS, and client-specific formats.
• Implement sim-to-real transfer tooling: domain randomization configs, Cosmos Transfer integration, and quality validation ensuring synthetic data improves policy performance.
• Build annotation platform backend systems — task taxonomy APIs, temporal segmentation, quality scoring, and inter-annotator agreement for robotics episode data.
• Integrate with NVIDIA Isaac ecosystem (Isaac Sim, Isaac Lab, GR00T-Mimic, OSMO) to orchestrate synthetic data generation on cloud GPU infrastructure.

What This Job Offers

• Job Type: Full-time
• Career Level: Mid Level