A gym is a flight simulator for an AI agent. It gives the model real tools, throws realistic curveballs at it, and runs a strict test script to check if the job was actually done right. Pass the test, get a reward. Fail the test, learn why and try again. Every gym below runs as an OpenEnv environment across 10 engineering verticals.
Navigate photorealistic multi-lane urban and highway environments in CARLA. Closed-loop sensor streams, trajectory tracking, collision avoidance, and traffic rule conformance gate reward.
OPENENV / CARLAManipulate articulated objects (doors, drawers, faucets, gears) using contact-rich physics in ManiSkill. Sensor feedback, joint torque limits, and multi-stage goal states determine success.
OPENENV / MANISKILLTrain dexterous robotic manipulation policies across MetaWorld and LeRobot benchmark tasks. Multi-camera observations, force feedback, and terminal spatial tolerances gate the reward.
OPENENV / LEROBOTExecute 3D waypoint navigation, wind disturbance rejection, and agile obstacle avoidance for quadcopters. MAVLink telemetry, flight corridor boundaries, and soft landing criteria gate reward.
OPENENV / PX4Achieve dynamic gait stabilization and terrain traversal for bipedal and quadrupedal systems. Ground reaction forces, contact phase timing, and energy economy gate qualification.
OPENENV / MUJOCOTune physical simulator parameters so replayed robot trajectories match recorded hardware telemetry within strict tolerance, compensating for motor latency and backlash.
OPENENV / SIM-TO-REALControl physical robotic manipulators over distributed Viam APIs. End-effector positioning, workspace bounding boxes, collision avoidance, and gripper force sensing gate execution.
OPENENV / VIAMDrive autonomous mobile robots (AMR) through dynamic warehouse layouts. LiDAR point clouds, visual odometry, costmap planning, and zero-collision obstacle clearance determine reward.
OPENENV / GAZEBOCoordinate decentralized swarms of autonomous agents moving toward shared spatial formations. Control barrier functions (CBF) and inter-agent collision boundaries gate qualification.
OPENENV / ROBOTARIUMCoordinate two Cartesian grippers in automatic MuJoCo simulation: uncap a marker, extract an eraser or pad, pour clip proxies, and hand over a ruler. Server-owned physical state, bilateral contacts, force limits, and stable task completion determine reward.
OPENENV / MUJOCORun native MuJoCo 3.11.0 / MjBatch 0.1.0 CPU batches with per-replica friction and actuator controls. Evaluate Go1 velocity tracking with a released PPO actor, G1 backflip MPC, arm friction identification from encoder recordings, and throwing-arm co-design.
NATIVE MUJOCO / MJBATCH