Train
Spin up a simulation environment and train an adaptive policy across 16+ reinforcement learning algorithms, from on-policy methods like PPO to offline methods like CQL, behind one dispatcher.
Adaptive learning platform
RLX is OptRL's platform for building adaptive, decision-making systems. It trains, evaluates, and deploys adaptive policies, with reinforcement learning as the technique underneath and explainable benchmarking built in. Every OptRL Benchmark entry below, and every system behind a client engagement, runs through RLX first. The leaderboard is proof of what it produces, not the product itself. Request access to run RLX on your own problem, or jump straight to the board.
Browse the leaderboard: published research ranked next to OptRL's own RLX-trained policies and Playground runs, all scored the same way.
RLX builds and runs every adaptive system on this page. See what it actually does before you dig into scores.
Open the Playground, pick an algorithm and an environment, and publish a run back to the leaderboard. No installs, no code.
The platform
Training an adaptive system, scoring it fairly, and shipping it behind a monitored API used to take three different toolchains. RLX is OptRL's single platform for all three, plus the explainability layer that makes the result auditable instead of a black box.
Spin up a simulation environment and train an adaptive policy across 16+ reinforcement learning algorithms, from on-policy methods like PPO to offline methods like CQL, behind one dispatcher.
Score every policy the same way: mean reward, stability, success rate, and OptRL's own sim-to-real gap, so results are comparable instead of cherry-picked.
Ship an adaptive policy behind a monitored API with a rollback path built in, the same pipeline behind every OptRL client engagement.
Attach reward curves, drift charts, and governance metrics to every policy automatically, so a result is auditable the day it ships, not reconstructed after the fact.
The leaderboard and the Playground are both built on RLX. What you can't do yet as a visitor is open RLX on your own problem, that part is invite-only for now.
The average score an algorithm earns per run in a given environment. Higher is better, but only compare mean reward within the same environment, since reward scales differ across tasks.
How consistent an algorithm is across repeated runs with different random seeds. A score near 1.0 means it lands in roughly the same place every time. A lower score means results swing more from run to run.
The share of episodes where the agent actually completed the task's goal, separate from how much reward it collected along the way.
How many environment interactions, or gradient steps for offline methods, the algorithm used to reach its reported score. A rough proxy for how expensive the run was.
An OptRL-specific score that measures how much performance drops when an algorithm faces scenarios outside its normal training distribution. Smaller is better.
The family an algorithm belongs to. On-policy methods like PPO learn only from data collected under the current policy. Off-policy methods like SAC reuse older data from a replay buffer. Offline RL methods like CQL learn from a fixed dataset with no live interaction. Imitation and inverse RL methods like GAIL learn from expert demonstrations instead of a reward signal.
Generate a verified result
OptRL Playground is a separate, interactive product. Pick an algorithm, run it on a real operational environment, and publish your run back here. No installs, no code required.