2026 Fall Robotics Colloquium: Rose Hendrix, Maya Guru, and Rohun Tripathi (AI2)

In this talk

1. 2:17

   Why robotics needs a data-generation engine

   Hendrix describes a future robot policy that can follow language, handle long tasks, adapt across environments, and improve with limited feedback. She argues that robot hardware is changing quickly, so datasets collected for today’s platforms may lose relevance. Her proposed investment is a repeatable way to generate data, with simulation complementing real-world teleoperation.

2. 8:11

   Lessons from synthetic-data research

   Hendrix traces AI2’s progression from SPoC and Polyformer to work on manipulation and cross-embodiment policies. She says procedural variety can help policies ignore irrelevant differences, while frozen visual encoders can reduce overfitting to simulated images. She also emphasizes broad randomization, modeling retries, and building data-generation tools that can be run again as methods change.

3. 12:53

   RoboSpaces and the MoMoBot data pipeline

   Hendrix and Guru describe RoboSpaces, an ecosystem with 230,000 scenes, 130,000 simulation-ready objects, and support for simulators including MuJoCo and Isaac. Tripathi explains how MoMoBot creates pick-and-place demonstrations: it samples scenes and objects, randomizes conditions, and uses a scripted simulator-aware demonstrator with action noise and retries. The saved data includes camera views, actions, robot state, and simulator-only information.

4. 19:12

   Policy results, limits, and next steps

   Tripathi presents MoMoBot policy variants trained on simulation data for Franka and RBW1 robots, then evaluated in simulation and across four real-world environments. The team reports gains from more trajectories and shows adaptations to changed table heights and camera positions, along with recovery after a lost grasp. The work focuses on rigid, mostly prehensile tasks and RGB plus robot-state inputs; future directions include broader tasks, deformable objects, and further use of depth.

5. 33:57

   Open questions about data and architecture

   In discussion, the team clarifies that some comparisons isolate data while others isolate architecture, but the evaluations do not complete every combination of data and model in a full comparison grid. The speakers say they see the data-generation engine as a central contribution, while noting that stronger comparisons are constrained by unavailable baseline datasets. They also discuss recovery training, frozen vision encoders, and using planners or goal poses to expand task variety.