RL Environments
Shipped games are the most tested, tuned environments ever built. We instrument them for RL: multiplayer tasks with humans and NPCs, graded by code, with custom capture on every run.
Verifier-graded tasks: labeling, repair, retargeting, simplification. Pass or fail, reason attached.
Most RL environments are built for the benchmark. These were built for millions of users to play, test, and game.
A shipped game survived years of balancing, QA and millions of hours of play. The physics hold up. The economy works. The NPCs fight back. That tuning is the hard part, and it comes with the license.
Tuned by shipping
Difficulty curves and physics that survived real players.
Multiplayer, properly
Human opponents, NPC teammates, full squads. Cooperative, adversarial and mixed.
Graded by code
Programmatic verifiers score every run. Pass or fail, with the reason attached.
Capture built in
Every rollout can return a configured selection of the 25 modalities. Training data comes back with the result.
The verifier sees what the eye misses
Your verifier here
Every environment takes custom checks: state assertions, thresholds, time limits, contact rules. If your trainer can read a scalar, it can run here.
Same task. Turn the dial.
One rover, reassembled from loose parts by a frontier coding agent, on three settings. Hand it 5 parts and it passes. 25, it places 6. All 62, it works for over an hour and places 23. The grader is never shown to the agent, and reward is binary.
From navigation to a working economy
Navigation & search
Find, reach, escape. Dynamic worlds, moving obstacles, changing light.
Combat & control
Micro against tuned NPCs, squad coordination, hold-the-point objectives.
Crafting chains
Gather, refine, build. Long-horizon plans with real dependency trees.
Economy & trade
Buy, sell, price. Markets that already balance because players tested them.
Cooperative play
Escort, rescue, defend. Agents with human or NPC teammates.
Deception & pursuit
Hide and seek against adaptive opponents, in worlds with real cover.
Interface mastery
Menus, inventories, maps. The UI skills agents need everywhere else.
Custom briefs
Your task, our worlds. Custom instructions, custom verifiers, custom capture.
Every task, fully specified
| Field | Example | Format |
|---|---|---|
| Goal | label every part of the vehicle | natural language plus schema |
| Verifier | part-label IoU at or above 0.60 | programmatic, deterministic |
| Reward | pass/fail, plus shaped variants | scalar per episode |
| Players | solo, human-vs-agent, mixed squads | configurable per run |
| Capture | selected modalities, on request | returned with the run |