Our team built Spice Terminator at the London edition of the Hugging Face LeRobot Worldwide Hackathon in June 2025. In 24 hours, we connected a learned robot policy, physical hardware and a browser interface into a working plate-cleaning demo.
From demonstrations to physical actions
We used an SO-101 arm equipped with a RealSense depth camera and an RGB camera. The task was simple to describe: find the sponge, pick it up, scrub a dirty plate and return the sponge. Making that sequence work meant coordinating perception, grasping and movement.
We collected teleoperated plate-cleaning demonstrations and fine-tuned NVIDIA GR00T N1.5 and Hugging Face SmolVLA. The resulting demo completed the cleaning sequence without human intervention.
24-hour build · top 30 global winner
A web interface for the robot
Alongside the cleaning pipeline, we built a browser platform to monitor and control the robot. A real-time 3D view kept joint positions synchronized with the physical arm, while camera feeds and telemetry brought the robot’s state into one interface.
The platform supported text and voice prompts to trigger the trained behavior, and control in either the simulation or the physical environment. We built on Haixuan Xavier Tao’s open-source work to get the platform running within the hackathon.
Watch the submission
Global recognition
Spice Terminator made the global top 30 winners list and won the Best Novelty Prize at the London event. NVIDIA Robotics later highlighted our team among the worldwide winners using GR00T.