About
Most RL post-training work assumes the environment lives somewhere outside the model. Here, the physics simulator runs natively on GPU – which changes the data flow, the optimisation surface, and what's actually possible. The team is using that to push RL post-training for robotics in directions that aren't available to labs working with standard LLM setups.
This is a well-funded physical AI lab building universal foundation models for general-purpose robots. The research stack spans both VLA models — mapping perception and language to robot actions — and world models that learn and predict environment dynamics. Both are in active development, and RL sits at the centre of how they're advancing both.
The team is small, deeply research-driven, and recruited from top AI labs.
What you'll do
- Design and run RL post-training pipelines for large multimodal robotics models, across both VLA and world model architectures
- Work with a GPU-native simulation environment, enabling data flow optimisations not available in standard LLM RL setups
- Apply sampling during training — across diffusion and autoregressive approaches — to avoid distribution drift
- Design reward models and feedback signals that transfer reliably from simulation to real-world robotic behaviour
- Iterate on both diffusion-based and autoregressive action models, applying RL where it pushes capability furthest
- Collaborate closely with training infra and simulation teams to turn research ideas into production-ready systems
- Run rigorous experiments and translate results into better models, not just better benchmarks
What you'll need
- Strong RL research background, applied to large-scale or real-world systems — LLM-background RL considered if the research track is strong
- Experience post-training large models using RL methods: RLHF, GRPO, PPO, or similar
- Proficiency in Python and PyTorch; experience with large-scale distributed training
- Track record from a top AI lab, frontier model team, or leading robotics or embodied AI group
- Publications at top-tier venues strongly preferred (NeurIPS, ICML, ICLR, CoRL, RSS, or similar)
Optional Bonus
- Experience with diffusion model RL or autoregressive action model training
- Background in sim-to-real transfer or embodied AI
- Familiarity with robotics control or physical data pipelines
Shortlisted candidates will be contacted within 48 hours.
