SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
SpAItial is building next-generation world models that natively understand physics and geometry, enabling applications across robotics, AR/VR, gaming, and cinema. This Research Scientist role focuses on training the policies that transform world models into acting robots.
You will own the complete vision-language-action (VLA) and world-action models (WAM) pipeline end-to-end: data curation, backbone selection, action representation design, training execution, and evaluation methodology. Your core mission is ensuring policies are genuinely competent in real-world deployment, not merely lucky in simulation.
Key responsibilities include:
- Owning the full training pipeline from raw data through deployed robot policy
- Contributing to technical direction for embodied AI research
- Closing the sim-to-real gap via domain randomization, system identification, and calibration
- Adapting vision-language model backbones for control tasks (encoder selection, adapter strategies, co-training)
- Curating and weighting training data across heterogeneous robot datasets with varying embodiments, action spaces, and sensors
- Designing action representations (tokenization, chunking, diffusion, flow-matching experts)
- Building world-model components that predict future observations conditioned on actions
- Executing post-training: supervised fine-tuning for target embodiments and RL for robustness
You bring a PhD in robotics, machine learning, or computer vision with demonstrated robot learning focus. You have publications at top venues (CoRL, RSS, ICRA, IROS, CVPR, ICCV, ECCV, NeurIPS), open-source contributions, or deployed systems. You have deep hands-on experience training modern robot policies (VLA, WAM, diffusion) end-to-end rather than fine-tuning released checkpoints. You understand imitation learning fundamentals and RL fine-tuning of pretrained policies. You are fluent with VLM backbones and their adaptation for control. You code expertly in Python and PyTorch with multi-node distributed training experience (FSDP or equivalent).
