This robot learning role is with a seriously exciting scale up. The platform is mature, the data is flowing, and the team is ready to scale its most promising research directions into production-grade manipulation policies.
They need someone to lead the development and deployment of large behaviour models, taking diffusion transformers, VLAs, and language-conditioned policies from the literature onto a real bi-manual humanoid.
This is not a research-only role. You'll inherit a mature policy training codebase, a VR teleoperation pipeline producing high-frequency multi-modal data, and a Gymnasium environment wrapping a real robot. The work you ship runs on hardware.
The Role
You will architect, train, and deploy end-to-end large behaviour models for bi-manual and mobile manipulation, and lead the maturing of the early-stage RL pipeline.
The key responsibilities
- Architect, train, and evaluate end-to-end large behaviour models for bi-manual and mobile manipulation
- Advance diffusion transformer policies, mature VLA integration, and develop language conditioning for true multi-task generalisation
- Apply RL to refine pre-trained policies: RL token fine-tuning, residual RL, off-policy RL with reference-action regularisation, RL-based fine-tuning of diffusion policies
- Build a systematic sim-to-real transfer pipeline, connecting existing simulation infrastructure to training
- Deploy and iterate learned policies on physical robot hardware
- Mentor junior researchers and engineers, and publish at top-tier venues
What We're Looking For
Essential:
- PhD/MSc in ML, Robotics, CS, or related field with 4+ years of equivalent industry research experience
- Demonstrated expertise training and deploying learned manipulation policies on real robots
- Strong background in at least two of: behaviour cloning, diffusion policies, VLA/VLM architectures, RL for manipulation
- PyTorch and large-scale (multi-GPU, distributed) training
- Track record of publications at top-tier venues (CoRL, RSS, ICRA, NeurIPS, ICML, ICLR), or equivalent demonstrated research impact through deployed systems, patents, or significant open-source contributions
- Strong Python; production-quality research code with proper testing, type hints, and documentation
Useful:
- Hands-on experience with humanoid or bi-manual manipulation platforms
- Diffusion transformer, ACT, or VLA architectures specifically
- Pre-trained vision/language models for robot control (CLIP, DINOv2, PaliGemma)
- MuJoCo, Isaac Sim, or ManiSkill for sim-to-real policy training
- RL fine-tuning of pre-trained policies (residual RL, DPPO, or similar)
- 3D perception for policy conditioning (point clouds, keypoints, NeRFs)
Key contribution areas
Policy Architecture & Training
- End-to-end large behaviour models for bi-manual and mobile manipulation
- Scale and evolve diffusion transformer policies, VLA integration, and language conditioning
- Extend the imitation learning pipeline to leverage growing teleoperation datasets
- Apply RL to push beyond what imitation alone can reach
- Target sub-millimetre precision and contact-rich manipulation
Generalisation & Scaling
- Develop policies that generalise across tasks, object categories, and environments
- Move from single-task to multi-task and task-conditioned architectures
- Design hierarchical behaviour systems for long-horizon manipulation
- Investigate data-efficient learning: few-shot adaptation, transfer learning, multi-dataset training
- Drive systematic ablations across architectures
Sim-to-Real & Deployment
- Build the sim-to-real transfer pipeline: domain randomisation, rendering augmentation, sim-to-real benchmarking
- Deploy and iterate learned policies on physical robot hardware
- Extend the Gymnasium environment wrapper and integrate with the robot's control stack
- Leverage perception team outputs (keypoints, learned features, 3D point clouds) for policy conditioning
Research Leadership
- Track the literature and bring relevant advances back to the team
- Identify and propose new research directions aligned with the manipulation roadmap
- Mentor junior researchers and engineers
- Publish at top-tier venues — conference attendance and open-source contributions are actively supported
What's On Offer
- Join a team with world class applied research scientists, ML engineers, and robotics software engineers
- A mature platform that ships to physical hardware, not slides
- Active support for conference attendance and open-source contributions
- Competitive compensation
Apply or send your CV to — Imogen@waverecruitment.co.uk