> Markdown version of [/jobs/ext/3721598-research-scientist-robot-learning-vla-wam](https://www.wearedevelopers.com/jobs/ext/3721598-research-scientist-robot-learning-vla-wam). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md).

---

# Research Scientist - Robot Learning (VLA / WAM)

- **Company:** SpAltial Ltd
- **Location:** UK
- **Experience:** Expert
- **Contract:** Permanent contract
- **Skills:** Computer Vision, Distributed Computing Environment, Python (Programming Language), Machine Learning, Open Source Technology, Pytorch, Decoding
- **Published:** October 14, 2026
- **Apply:** https://www.totaljobs.com/job/research-scientist/spaltial-ltd-job108163718

## About the Role

* A PhD in robotics, machine learning, or computer vision with a robot learning focus, from the PhD alone or followed by industry experience.
 * Publications at top venues such as (CoRL, RSS, ICRA, IROS or CVPR, ICCV, ECCV, NeurIPS), open-source work, and/or deployed systems.
 * Deep experience with modern robot policy designs (VLA, WAM, diffusion), trained end to end rather than fine-tuned from a released checkpoint.
 * Strong imitation learning fundamentals, and familiarity with RL fine-tuning of pretrained policies.
 * Fluency with VLM backbones and how to adapt them for control.
 * Expert Python and PyTorch, with multi-node distributed training experience (FSDP or equivalent).

## Description

* Own the training pipeline for vision-language-action (VLA) and world-action models (WAM) end to end, from data to a policy running on a robot.
 * Contribute to setting the technical direction for embodied research at SpAItial.
 * Close the sim-to-real gap through domain randomization, system identification, and calibration, and build evaluation that predicts real-world transfer.
 * Adapt VLM backbones for control: encoder choice and adapter strategies, co-training.
 * Curate and weight the training mix across heterogeneous robot datasets, spanning differing embodiments, action spaces, and sensor setups.
 * Design action representation and decoding, including tokenization, chunking, diffusion, and flow-matching action experts.
 * Build the world-model components that predict future observations conditioned on action.
 * Run post-training: supervised fine-tuning onto target embodiments, and RL for robustness beyond demonstrations.