The RoleWe are looking for a
Research Scientist, RL & Simulation to own the RL + simulation engine that turns large-scale human demonstrations into scalable robot learning signals.
This is a research-meets-systems role: you'll build simulation environments, retarget human motion to robot actions, train and evaluate policies, and drive sim-to-real transfer with clear metrics.
What You'll Work OnSimulation Environments- Build and maintain simulation environments for robotics learning (e.g., Isaac Sim / Isaac Gym, MuJoCo, Genesis, Habitat, ManiSkill).
- Decide what environments and assets to build first to maximize learning velocity.
Retargeting (Human 12 Robot)- Convert human demonstrations into robot-executable trajectories.
- Explore IK-based, optimization-based, and learning-based retargeting approaches.
Policy Learning & Evaluation- Train policies from demonstrations using imitation learning + RL:
- Behavior Cloning, DAgger-style aggregation, Offline RL
- PPO / SAC (or similar) when online fine-tuning is required
- Define evaluation: success metrics, stress tests, generalization, and regression tracking.
Sim-to-Real- Drive transfer via domain randomization, system identification, contact modeling, and failure-mode analysis.
- Use real data to identify domain gaps that matter.
Who You AreRequired Background- MSc/PhD (or equivalent research experience) in robotics, ML, or a related field.
- Strong hands-on experience with robot simulation and policy learning.
- Proficiency in Python; solid engineering discipline (reproducible experiments, clean code, debugging).
- Comfort working end-to-end: environment 12 data 12 training 12 evaluation.
Strong Signals:- Experience with manipulation, dexterous hands, or locomotion.
- Experience with retargeting, IK, trajectory optimization, or differentiable simulation.
- Deep intuition for what makes sim-to-real succeed or fail.
Why This Role- Define how Mecka turns egocentric human behavior into scalable robot learning signals.
- High ownership, fast iteration, and direct connection to real-world datasets.
