Research Scientist Intern

You'll work end-to-end: from training policies in simulation all the way to debugging what breaks when the policy meets real hardware. The work spans simulation fidelity, policy learning, and empirical bring-up on muscle-driven robots. Policy training & simulation Train RL policies in simulation — MuJoCo, Isaac, or similar — for dexterous, in-hand manipulation on musculotendon-driven robots. Improve the fidelity of our simulation models of compliant, high-DOF musculoskeletal systems : the closer the sim, the smaller the reality gap. Design metrics and benchmarks to evaluate policies in open loop, in simulation, and on the real robot — so progress is measurable at every stage. Build and harden sim-to-real pipelines : domain randomization, system identification, and actuator modeling for MTUs and tendon routing. Deploy policies on real hardware, then debug what breaks : latency, friction, hysteresis, sensing, and everything the simulator didn't warn you about. Requirements We care about fundamentals and hands-on experience with real systems. Strong theoretical grounding matters, but so does comfort operating outside the simulator. Required Currently pursuing or recently completed a BS, MS, or PhD in CS, Robotics, EE, ML, or a related field. Strong fundamentals in reinforcement learning — e.g. PPO, SAC — and deep learning, with hands-on PyTorch or JAX . Experience training RL policies , ideally for robotics or continuous control. Comfort with a physics simulator : MuJoCo, Isaac Sim / Lab, Brax, Genesis, or similar. Willingness to work with real hardware — robotics is an empirical science, and debugging a policy on real robot is part of the job. Sim-to-real transfer, domain randomization, or system identification. Dexterous manipulation, contact-rich control, or tendon-driven systems. Published work at ICRA, IROS, CoRL, RSS, ICML, NeurIPS, CVPR, ECCV, ICCV, or similar. What We Offer You'll be contributing to an open research problem with direct impact on the robot we're building at Clone, on a path to realize the most human-like and human-level android in the world. End-to-end ownership. From simulation to real hardware deployment — you'll own the full stack for your track, not just the training loop. Hardware access. Direct access to muscle-driven robotic hands and the full sensor stack — not a simulation-only role. Mountain View lab. On-site in the Bay Area, alongside the core Intelligence & Behavior and Demos teams. Compute. The GPU resources to train the policies the work actually requires. Project guidance. Close collaboration with researchers who care about sim-to-real, musculoskeletal systems, and getting things to work on real hardware. Scaling. Our company is constantly growing. You will become part of an international team with wide development opportunities. 3-month internship (including 8 days of paid holidays) with the possibility to extend to a full-time job. #J-18808-Ljbffr