Want to know if this job is worth applying to?
San Francisco, California, United States
Onsite
Maven Robotics is building the world’s leading general-purpose robots and providing physical AI solutions for the most challenging industrial autonomy tasks.
Operating in stealth, we are assembling a team of world-class innovators who think from first principles. Our mission is to achieve human-level task success rates in complex environments, even when faced with limited fine-tuning data or evolving robotic hardware. We value unwavering truth-seeking, humility, and relentless determination.
We are seeking exceptional AI researchers and engineers to architect the neural backbone of our general-purpose robots. You will design the Vision-Language-Action (VLA) frameworks or Action World Models that allow our robots to reason, adapt, and succeed where traditional automation fails.
Note on Leveling: We are hiring across all levels (from early-career to staff/principal). You are not expected to be a master of every domain listed below; however, you must be able to justify world-class excellence in at least one core factor (e.g., model architecture, RL formulations, or high-scale data systems).
In this role, you will:
Architect Embodied Foundations: Design model architectures (VLA, World Models, etc.) that achieve ultra-high task success rates with minimal human demonstrations.
Master Data Efficiency: Develop novel co-training strategies and efficient learning algorithms that leverage diverse data sources—from Internet-scale video to sparse, high-fidelity human interventions.
Generalize Cross Embodiments: Build models capable of zero-shot or few-shot adaptation to new robot configurations, maintaining a high success rate even when proprietary hardware and actuation systems evolve.
Innovate Real-World RL: Formulate and deploy novel Reinforcement Learning and policy extraction methods specifically designed for physical, real-world manipulation.
Design the Data Loop: Collaborate on advanced data collection systems to capture critical human intervention data for model bootstrapping.
Must-have:
MS or PhD in CS, Robotics, Machine Learning, or a related field (or equivalent industry experience).
Deep Technical Mastery: Advanced understanding of transformers, multi-modal alignment, and mapping perception to high-frequency motor control.
Specialized Excellence: Proven ability to innovate—not just implement—within one or more areas: VLA models, Real-world RL, or large-scale Data Infrastructure.
Software Excellence: Expert-level Python and deep familiarity with PyTorch or JAX
Nice-to-have:
A track record of high-impact publications (NeurIPS, ICRA, RSS, CVPR) or significant open-source contributions.
Experience with large-scale distributed training and model compression.
General knowledge of robotics principles (kinematics, dynamics).
You will be working on a unique in-house robot platform that is evolving alongside your models. We don't just want to reach 90% success; we are building the architectural breakthroughs required to reach 99.9% in the real world.