Real-to-Sim Reconstruction: Build a pipeline that reconstructs real, facility-scale environments into simulation (3D reconstruction, photorealistic re-rendering). That means full rooms and aisles our wheeled base navigates, not just tabletop scenes.
Facility-Scale Scene Generation: Procedurally generate navigable environments (layouts, obstacles, object placement) that stress-test loco-manipulation policies across the range of spaces our robots actually operate in.
Loco-Manipulation Sim-for-Data-Gen: Generate synthetic training data for policies that jointly reason about base positioning and arm manipulation (approach angles, reachability, obstacle-aware repositioning), and feed real-world failure modes back into new sim scenarios.
Loco-Manipulation Evaluation: Build simulation benchmarks that test the joint base+arm policy (VLA / imitation learning / diffusion models) as a whole, not manipulation in isolation on a fixed base.
Rendering for ML, Not Just Physics: Push photorealistic rendering quality specifically to close the sim-to-real gap for vision-based loco-manipulation. Rendering fidelity is a first-class deliverable, not an afterthought on top of physics accuracy.
Cross-Team Collaboration: Partner with the AI Research team on where simulated data and evaluation most accelerate loco-manipulation policy development, and with Data Ops on what’s worth capturing in the real world vs. generating in sim.