David J. Watkins

David J. Watkins

Head of Research @ Tutor Intelligence | Foundation models and robot manipulation at scale | Advisor @ Tola Capital | Built teams of 40 | Ph.D. Columbia

David Watkins - Head of Research

I’m Head of Research at Tutor Intelligence, where I lead robot learning research connecting policy training, demonstration quality, and deployment on a 100-robot bimanual fleet. I also advise Tola Capital on robotics and AI investments. Previously I was a Research Lead at the RAI Institute (formerly Boston Dynamics AI Institute), where I built the Foundation Models and Capture teams, foundation models, and large-scale data infrastructure for robot learning.

Research at Tutor Intelligence

  • Led team improvements to full-rate, synchronized data logging and headless teleoperation; co-authored DF1: Maintaining Demonstration Quality in a 100-Robot Teleoperation Pipeline (RSS 2026 Workshop: It’s the Demos)
  • Moved research training to a 32-H100 allocation and experiment tracking to self-hosted MLflow
  • Designed Rust middleware for data collection and policy execution, and brought an initial version onto robot hardware with motion inhibited
  • Set the research roadmap for policy evaluation and learning from human interventions

Selected Research

  • BRIDGE combines handheld demonstrations with targeted teleoperation through state-gated diffusion policy experts for contact-rich manipulation. Under review at ICRA 2027.
  • Koala Gripper co-designs handheld and actuated grippers with matched geometry, sensing, and force transmission for dexterous manipulation learning. Under review at IEEE Transactions on Robotics (T-RO).
  • DF1 connects teleoperation, data curation, and policy training with human interventions to maintain demonstration quality at fleet scale. RSS 2026 Workshop: It’s the Demos.

Foundation Models & Large-Scale Training

  • Built foundation models research capability from scratch; pitched, launched, and led a team of 10+ researchers
  • Designed and trained video prediction models using internet-scale video data on a 320 H100 GPU cluster (early 2023)
  • Created Theia, a vision foundation model that distills multiple pretrained models into a single compact representation (CoRL 2024)
  • Built multimodal architectures combining diverse sensor modalities with internet-scale pretrained priors for improved downstream performance

Data Infrastructure & Quality at Scale

  • Founded and led the Capture team (30 people), the largest data collection effort at the institute with a roadmap for 100,000+ demonstrations
  • Built handheld force-based data collection capturing force, vision, and proprioception for in-the-wild demonstrations
  • Created task definition frameworks and benchmark protocols that improved demonstration quality and consistency across teams
  • Established research partnerships with Google, Columbia, ETH Zurich, and Agile Robots for data and evaluation

Reinforcement Learning from Human Feedback

  • Won 1st place at the MineRL BASALT Competition (NeurIPS 2021) for learning from human feedback, combining imitation learning, human preference data, and hierarchical knowledge engineering to solve tasks defined only by natural language descriptions
  • Developed DIP-RL, a preference-based RL algorithm that leverages demonstrations to infer reward functions from human preferences (ICML 2023 Workshop)
  • Developed gradient-free RL enabling online learning with non-differentiable semantic reward functions (U.S. Patent pending, May 2025)

Publications & Writing

Education & Background

More information is available in my curriculum vitae or my resume.