Researcher, World Models

What You'll Do

• Design, train, and rigorously evaluate world models that let Asimov predict the consequences of actions across visual, proprioceptive, and force/torque modalities.
• Advance our self-supervised learning stack for visual and sensor representations, building on the JEPA family (V-JEPA, I-JEPA, and related predictive-embedding approaches).
• Prototype and benchmark generative and predictive architectures (diffusion, DiT, flow matching, VAEs) against JEPA-style objectives for embodied prediction and planning.
• Own the data pipeline for your experiments end to end: curation, tooling, and scaling.
• Integrate what you build with our platform, firmware, and software teams so research reaches the robot.
• Contribute to sim-to-real transfer, inverse dynamics, and multi-modal sensor fusion, and publish or open-source work where it strengthens the field.

What We Look For

• Proven modeling track record: trained models with solid, honest evaluations.
• JEPA fluency: understanding of joint-embedding predictive approach and its alternatives.
• Breadth across approaches: familiarity with VLA models and trade-offs.
• Depth in a modality: strong depth in vision, audio, natural language, or similar.
• Strong data abilities: ability to work without a data-engineering team.
• Solid engineering: ability to implement, integrate, and ship alongside teams.
• Conversant, ideally deep, in several of: SSL for visual/sensor representations; world models (JEPA variants); generative/predictive architectures; robotics ML (VLA, inverse dynamics, sim-to-real, optical flow); sensor fusion; PyTorch, JAX, distributed training.

Nice to Have

• Publications at NeurIPS, ICML, ICLR, CoRL, RSS (or arXiv with comparable citations).
• PhD or equivalent research experience in ML, robotics, or computer vision (not required with strong portfolio).
• Demonstrated hardware or robotics interest or hands-on experience.
• Strong communication: technical blogs, talks, or clear written research.