Zero-Shot Policy Comparison of DreamZero and π0.5
DreamZero is a World Action Model by NVIDIA that replaces the Vision-Language backbone common in VLAs with a video diffusion model to inherit richer physical priors. I compare DreamZero with π0.5 across 12 tasks in Isaac Sim on a DROID setup.
Two-Legged Spot: Teaching a Quadruped to Only Use Its Hind Legs for Locomotion
We use reinforcement learning in NVIDIA Isaac Lab to train Boston Dynamics’ Spot to move using only its hind legs. Despite several challenges, we deploy the policy in the real world and show how well the sim-to-real policy transfer works.
Grievous: A General-Purpose Household Robot
Grievous is a low-cost mobile manipulation platform built on XLeRobot, extended with onboard SO-101 leader arms for fast in-place teleoperation. We finetune SmolVLA on pick-and-place and dusting tasks collected in lab environment and showcase current progress.
Multi-Agent Decision-Support in Adversarial Environments
Military logistics planning under adversarial conditions is notoriously brittle. At Multi AI, we built a coordination platform integrating decentralized MCTS and generative AI to optimize theater-scale plans — delivering 38% more resources in field tests.
Autonomous Mission Planning for Unmanned Robotics Systems
From voice command to synchronized flip: we show our early autonomous mission planning work that grew into Multi AI, including Tello drone choreography and a loyal wingman simulation for unmanned-manned teaming.
3D Printing with Drones
What if a drone could build structures humans can’t reach? We investigate integrating underactuated multicopters with 3D printing for large-scale remote additive manufacturing — demonstrating aerial positioning accuracy competitive with ground-based robotic systems.
Flying in Cluttered Environments
Standard path planners proved insufficient for dense multi-drone environments. We develop custom algorithms on top of EGO-Planner to enable robust obstacle avoidance in the Texas Robotics motion capture space, supporting advanced multi-agent research.
Motion Capture Space with Submillimeter Position Sensing
35 × 35 × 20 feet of submillimeter-accurate airspace — built from the ground up. We designed, constructed, and commissioned the Texas Robotics motion capture facility at UT Austin (Anna Hiss Gymnasium) to serve as the backbone for robotics research.
First 3D Printing Attempt with a Drone
Unstable flight, premature extrusion, material that refuses to adhere — this is what aerial 3D printing looks like at the start. We document the first prototype test of a multi-year aerial additive manufacturing research program, including what went wrong and why.
