DeliveryGym: adaptive UE5 RL environment boosts Qwen3-VL-4B delivery earnings 54.3%
Lianhuiq · x · 2026-09-21
DeliveryGym is an adaptive RL environment built in Unreal Engine 5 where embodied agents navigate Paris, deliver food, and earn money. The key idea: earnings are the reward and failures shape the curriculum — the environment automatically finds where the agent struggles and generates harder tasks targeting those weaknesses, forming an embodied recursive self-improvement (RSI) loop. With this loop, Qwen3-VL-4B improves net income by 54.3% per shift.
Much of the work is systems-side: making long-horizon agentic online RL practical in UE5 while balancing photorealistic rendering with rollout throughput. The team uses SPEAR as the Python-to-Unreal interface, plus distributed rollout infrastructure with parallel simulator workers and asynchronous coordination across environment execution, inference, and policy training.
More from Embodied
- Defunct Eidon AI open-sources 13,451 egocentric videos with IMU data for robot learning — CyberRobooo · 2026-09-22
- Eidon AI Open-Sources 1,274-Hour Egocentric Human Activity Dataset for Robot Learning — CyberRobooo · 2026-09-22
- NVIDIA: scaling physical AI demands safety at every layer, from AVs to robots — nordicinst · 2026-09-22
- Apple wins the AI infra lottery: M5 Ultra packs 512GB unified memory for local AI — Hesamation · 2026-09-21
- Odyssey-3: one frozen world model drives humanoids, cars, drones, and games — thione · 2026-09-21
- Train a balance bot with PPO: Reinforcement Learning for Robotics Part 2 — ShawnHymel · 2026-09-21