Real world progress.
AI development has rapidly accelerated in the last decade, but far too often behind closed doors. We believe a world with abundant physical AI is positive-sum for all involved: safer, cleaner, and more prosperous. Therefore, we choose to share our advances publicly and help usher in a world where physical AI is responsibly deployed at scale. Much of our research focuses on the waste sector and autonomy, but the methods and insights are broadly applicable to other physical work. We openly share our results, benchmarks, and publications here, and welcome collaboration with other researchers and institutions.
Featured work
Result · August 2026Reinforcement learning outperforms a classical pipeline on a route-based service task
TrashBench v0 evaluates a unified policy for simulated residential collection. Across 100 held-out bin configurations, the learned policy reduces mean route time by 28.9% and off-pavement driving from 17.7% to 0.7% relative to a modular baseline.
Read the result →Benchmarks
TrashBench
A reinforcement-learning benchmark for curbside collection: a neighborhood loop, randomized bins, one policy for driving and manipulation, and an official 100-route evaluation protocol with published baselines. The environment suite, specification, and leaderboard are being prepared for open release so results can be reproduced, or beaten.
What's ahead
Future tiers of the benchmark add bins on both sides of the street, parked cars, more complex road networks, degraded sensing, and multi-vehicle coordination, each step closer to the streets we operate on. New results will be published here as they land.
Research is a team sport.
If teaching machines essential physical work is the kind of problem you want to spend your life on, we are hiring.
Join us