Artificial Intelligence Researcher

Auxo Talent; San Francisco (CA); On-site; USD 140,000 - 190,000; Full time

Job summary

Auxo Talent partners with an innovative robotics company developing Asimov, an open-source humanoid platform and the full software stack powering it. The Researcher – World Models will work at the intersection of self-supervised representation learning, predictive architectures and embodied control on real robots.

You will design, train and evaluate world models across multimodal data, advance the self-supervised stack, prototype predictive architectures, and own the data pipeline while

Qualifications

• Proven track record of training and evaluating ML models with rigorous evaluation.
• Experience with JEPA or similar predictive architectures.
• Strong data handling and experimentation pipeline skills.
• Experience in robotics or hardware interfacing.
• Publication history or arXiv work is a plus.
• Excellent communication of technical topics.

Responsibilities

• Design and evaluate world models for action prediction across visual, proprioceptive and force/torque modalities.
• Advance self-supervised learning stack for visual and sensor representations.
• Prototype generative and predictive architectures (diffusion, VAEs) against JEPA objectives.
• Own the data pipeline: data curation, tooling and scaling.
• Collaborate with platform/firmware/software teams for deployment to real robots.
• Publish or open-source research to strengthen the field and team goals.

Skills

JEPA

Self-supervised learning

World models

Data pipelines

PyTorch

JAX

Robotics ML

Education

PhD in ML/Robotics/Computer Vision

Tools

PyTorch

JAX

Python

Job description

We are partnering with an innovative company developing Asimov, an open-source humanoid robotics platform and the full software stack that powers it. Their mission is to make humanoid labour economically viable by turning advanced software into physical labour at scale.

Researcher – World Models

You will work at the intersection of self-supervised representation learning, predictive architectures and embodied control, collaborating closely with platform, firmware and hardware teams to bring research directly onto the robot.

The Role:

• Design, train and rigorously evaluate world models that enable Asimov to predict the consequences of actions across visual, proprioceptive and force/torque modalities.
• Advance the company's self-supervised learning stack for visual and sensor representations, building on and extending the JEPA family, including V-JEPA, I-JEPA and related predictive-embedding approaches.
• Prototype and benchmark generative and predictive architectures, including diffusion models, DiT, flow matching and VAEs, against JEPA-style objectives for embodied prediction and planning.
• Own the data pipeline for your experiments end-to-end, including data curation, tooling and scaling, without relying on a separate data engineering team.
• Integrate your research with platform, firmware and software teams to ensure your work reaches the robot rather than remaining purely experimental or theoretical.
• Contribute to sim-to-real transfer, inverse dynamics and multimodal sensor fusion.
• Publish or open-source research where it strengthens the field and contributes to the team's goals.

Qualifications:

They're looking for someone with a proven track record of training and evaluating machine learning models, with a focus on rigorous evaluation rather than simply optimising training metrics.

• Strong understanding of JEPA and joint-embedding predictive approaches, including where they fit relative to alternative architectures.
• Broad knowledge of adjacent approaches, including VLA (Vision-Language-Action) models, and an understanding of their respective trade-offs.
• Deep expertise in at least one sensory modality, such as computer vision, audio or natural language.
• Strong data skills and the ability to independently build, curate and scale datasets and experimentation pipelines.
• Strong software engineering skills, with the ability to implement, integrate and deploy research alongside platform, firmware and software teams.
• Experience across several of the following areas:
• Self-supervised learning for visual and sensor representations
• World models, including JEPA, V-JEPA, I-JEPA, LeJEPA and MJEPA
• Generative and predictive architectures, including diffusion, DiT, flow matching and VAEs
• Robotics ML, including VLA, inverse dynamics, sim-to-real and optical flow
• Sensor fusion across vision, proprioception and force/torque
• PyTorch and/or JAX

Advantageous Qualifications:

• Publications at leading conferences such as NeurIPS, ICML, ICLR, CoRL or RSS, or comparable research published on arXiv with demonstrated impact.
• PhD or equivalent research experience in machine learning, robotics or computer vision. A PhD is not required where there is a strong research portfolio.
• Demonstrated interest in or hands-on experience with robotics and physical hardware.
• Strong technical communication skills, demonstrated through technical blogs, talks or clear written research.