Evaluate
Run fixed training probes.
DeliveryGym: An RL Environment for Long-Horizon Embodied Agent Planning with Adaptive Curriculum
Adaptive RL environments towards embodied RSI.
1 UC San Diego2 Independent3 Johns Hopkins University4 Cornell University
Demo Overview
Deliver food under real-world constraints
Maximize delivery profit while respecting time, location, energy, money, food condition, and safety constraints.
The RL training loop
UE5 executes actions and verifies outcomes. Delivery profit and safety feedback provide the reward for online RL.
Online RL improves delivery performance
+54.3%
test net income
$7.60 → $11.73
Qwen3-VL-4B before and after safety-aware RL.
Adaptive environment for embodied RSI
The environment identifies the agent’s weaknesses and adapts the next round of training tasks.
Diagnose
Identify weak capabilities.
Generate
Generate targeted, feasible tasks.
Continue RL
Update the policy and repeat.
+16.5%
test net income
Adaptive curriculum vs. uniform sampling.
Same rollout budget and test suite.
A scalable environment for embodied RL
Scale across tasks and city maps, with 2,400 training trajectories per run.
available training configurations
+47% validation incometraining city maps
+61% validation incomeCitation