Robotics Simulation & Synthetic Data Intern

Robotics Simulation & Synthetic Data Intern - Physical AI Team

Location: Mountain View, CA Type: Internship (3–6 months) Team: Robotics / Physical AI 

About Deccan AI

Deccan AI is a model training and evaluation startup headquartered in Mountain View, CA. Founded by IIT Bombay, IIM Ahmedabad, and ex-Google alumni, we work with top AI frontier labs—including Google DeepMind and Snowflake—to build expert-curated datasets and evaluation infrastructure that power the next generation of AI. We are backed by Prosus Ventures, with a delivery center in Hyderabad, India.

Our Physical Intelligence practice is building the data backbone for embodied AI—synthetic trajectories, human demonstration datasets, and evaluation frameworks that help robots learn manipulation, locomotion, and reasoning skills at scale.

The Opportunity

Training the next generation of robot foundation models demands an order-of-magnitude more data than manual teleoperation can deliver. NVIDIA's Isaac GR00T and Cosmos pipelines have shown that a handful of human demonstrations can be amplified into hundreds of thousands of physically accurate synthetic trajectories—cutting months of data collection to hours.

As a Robotics Simulation & Synthetic Data Intern, you will own the end-to-end synthetic data generation pipeline: from configuring simulation environments in NVIDIA Isaac Sim and Isaac Lab, to running GR00T-Mimic trajectory expansion, to applying Cosmos Transfer for photorealistic domain augmentation. Your output will directly feed into training runs for VLA (Vision-Language-Action) models and be delivered to Deccan AI's frontier lab clients.

This is a high-impact, hands-on role. You won't be writing documentation or shadowing—you'll be generating the data that makes robots smarter.

What You'll Do

• Build and operate synthetic trajectory pipelines using NVIDIA Isaac Sim, Isaac Lab, and the GR00T-Mimic blueprint to generate large-scale manipulation and locomotion datasets from a small seed of human demonstrations.

• Configure simulation environments—scene composition, physics parameters, robot URDF/USD models, camera placements, and domain randomization—to maximize trajectory diversity and sim-to-real transfer quality.

• Apply world-foundation-model augmentation (NVIDIA Cosmos Transfer / GR00T-Dreams) to transform sim-rendered frames into photorealistic training images with varied lighting, textures, and backgrounds.

• Design and run data quality experiments—measure success rates, trajectory smoothness, and visual fidelity; iterate on pipeline parameters to improve downstream policy performance.

• Curate and version datasets in formats compatible with VLA model training (e.g., Open X-Embodiment, LeRobot), ensuring metadata, task labels, and action annotations meet client specifications.

• Benchmark synthetic vs. real data by training imitation-learning policies (e.g., ACT, Diffusion Policy) on mixed datasets and reporting sim-to-real transfer metrics.

• Document pipelines and author technical guides so the team can reproduce, scale, and extend your work beyond the internship.

What You'll Bring

Required

• Currently pursuing (or recently completed) an MS or PhD in Robotics, Computer Science, Mechanical Engineering, or a related field.

• Hands-on experience with at least one physics simulator: NVIDIA Isaac Sim/Isaac Lab, MuJoCo, PyBullet, or Gazebo.

• Strong Python skills and comfort working in Linux/Docker environments with GPU workloads (CUDA, multi-GPU scheduling).

• Familiarity with robot learning fundamentals: imitation learning, behavior cloning, reinforcement learning, or sim-to-real transfer.

• Working knowledge of USD/URDF scene descriptions and at least one robotics middleware (ROS/ROS2).

Preferred (any of these are a plus)

• Prior experience with NVIDIA Omniverse, Isaac Sim extensions, or Replicator for synthetic data generation.

• Familiarity with VLA architectures (RT-2, Octo, π0, GR00T N1) or open robot datasets (Open X-Embodiment, DROID, BridgeData).

• Experience with domain randomization, photorealistic rendering, or neural rendering (NeRF, Gaussian Splatting) for sim-to-real.

• Published research or course projects in robot manipulation, locomotion, or synthetic data for embodied AI.

Why This Internship

• Founding-stage impact. You're joining the robotics practice at inception. Your pipelines will define how Deccan AI generates physical AI data for years to come.

• Frontier lab exposure. Your work will be consumed by clients like Google DeepMind. Few internships put you this close to the cutting edge of embodied AI.

• Full-stack ownership. Sim environment setup, trajectory generation, visual augmentation, dataset curation, and policy benchmarking—you'll touch every layer.