LavenderSim alphaGitHub ↗
13 registered tasks · native renders

See the problem before you train it.

Every frame below comes from LavenderSim’s C++ RGB camera, using the same Python scene compiled by the RL environment. The gallery pairs visible mechanics with the actual policy interface.

Six LavenderSim environments rendered by the native camera
One engine, different control problems. Terrain locomotion, command-conditioned pushing, balance, navigation, and spatial manipulation all use the same scene/runtime boundary.
Mechanics auditedAll 13 environments pass seeded reset, bounded-space, finite rollout, termination, and native scene construction checks.
Truthful visualsRegenerate every image with .venv/bin/python docs-site/scripts/render_scenes.py.
Baseline statusThese are RL-ready tasks and smoke examples—not claims of benchmark-quality solved policies.

Choose by learning problem

If you want to learn…Start withThen try
Whether installation, vectorization, and PPO workCartPoleBalance-v0AcrobotSwingup-v0
Goal-conditioned spatial controlArmReach-v0PegInsertion-v0
Navigation from local sensingRoverWaypoint-v0TiltMaze-v0
Command-conditioned locomotionHopperCommand-v0QuadrupedTerrain-v0
Contact-rich manipulationTactileGripperLift-v0ManipulatorPush-v0
Control

Control environments

3 tasks
Native RGB camera render of CartPoleBalance-v0native render

CartPoleBalance-v0

1A / 12O

Push one cart while a passive pole responds to gravity. A compact check for effort control, activation dynamics, and snapshot replay.

effort actuatoractivationsnapshot
python examples/envs/cartpole_balance.py --steps 600 --web
Native RGB camera render of AcrobotSwingup-v0native render

AcrobotSwingup-v0

1A / 14O

Inject energy through one elbow motor and swing an underactuated two-link mechanism above its shoulder.

underactuationtip sitejoint friction
python examples/envs/acrobot_swingup.py --steps 600 --web
Native RGB camera render of ReactionWheelCube-v0native render

ReactionWheelCube-v0

3A / 24O

Stabilize a pivoted cube using three internal reaction wheels and imperfect inertial sensing.

delayed IMUarmature3-axis effort
python examples/envs/reaction_wheel_cube.py --steps 600 --web
Navigation

Navigation environments

2 tasks
Native RGB camera render of TiltMaze-v0native render

TiltMaze-v0

2A / 20O

Tilt a nested two-axis board, roll a low-friction ball to a target, and keep it stationary there.

2-axis gimballow frictiondwell success
python examples/envs/tilt_maze.py --steps 600 --web
Native RGB camera render of RoverWaypoint-v0native render

RoverWaypoint-v0

2A / 28O

Drive a differential rover across uneven terrain toward randomized local waypoints using a ray-fan scan.

ray fanheightfieldwheel velocity
python examples/envs/rover_waypoint.py --steps 600 --web
Locomotion

Locomotion environments

4 tasks
Native RGB camera render of HopperCommand-v0native render

HopperCommand-v0

3A / 28O

Track a requested planar speed with one articulated leg, a spring tendon, an IMU, and a ground probe.

fixed tendonIMUterrain probe
python examples/envs/hopper_command.py --steps 600 --web
Native RGB camera render of QuadrupedTerrain-v0native render

QuadrupedTerrain-v0

12A / 64O

Track randomized planar commands with twelve joints over a deterministic uneven heightfield.

12 DoFterrain raysheightfield
python examples/envs/quadruped_terrain.py --steps 600 --web
Native RGB camera render of WalkerCommand-v0native render

WalkerCommand-v0

6A / 25O

Crab-walk a six-joint biped to its body-relative right using only IMU and joint proprioception.

body-relative commandIMUcrab walk
python -m experiment.train_ppo --task WalkerCommand-v0
Native RGB camera render of QuadrupedCommand-v0native render

QuadrupedCommand-v0

12A / 37O

Track a browser-selected body-relative direction and speed on flat ground with the original quadruped task.

omnidirectionallive commandPPO
python -m experiment.train_ppo --task QuadrupedCommand-v0
Manipulation

Manipulation environments

4 tasks
Native RGB camera render of ArmReach-v0native render

ArmReach-v0

3A / 24O

Move a true spatial yaw–shoulder–elbow arm to randomized feasible 3-D points and settle its tip.

3-D kinematicstip velocitygoal site
python examples/envs/arm_reach.py --steps 600 --web
Native RGB camera render of TactileGripperLift-v0native render

TactileGripperLift-v0

2A / 30O

Close oppositely coupled jaws around a payload, lift it, and hold it with low slip.

tactile arraysjaw tendonwrist wrench
python examples/envs/tactile_gripper_lift.py --steps 600 --web
Native RGB camera render of PegInsertion-v0native render

PegInsertion-v0

3A / 36O

Move an arm-mounted peg toward a socket frame while balancing alignment, insertion depth, and contact wrench.

mounted toolforce/torquestable seating
python examples/envs/peg_insertion.py --steps 600 --web
Native RGB camera render of ManipulatorPush-v0native render

ManipulatorPush-v0

3A / 19O

Approach a ball from the useful side, push it to a clicked floor target, then leave it stationary.

goal-conditionedweb clicksettling
python -m experiment.train_ppo --task ManipulatorPush-v0

Create any registered task

python
from lavendersim.env import list_envs, make

for item in list_envs():
    print(item.id)

env = make("RoverWaypoint-v0", control_hz=60, horizon_seconds=3, seed=4)
observation, info = env.reset(seed=4)
print(env.observation_space.shape, env.action_space.shape)

What “verified” means here

The gallery is generated only after package tests exercise each registered ID. The suite checks repeated reset with the same seed, declared action and observation shapes, finite observations and rewards, invalid-action rejection, scene availability, and both success/failure metadata. Targeted tests also check the tilt gimbal hierarchy, opposing jaw tendon coefficients, spatial arm axes, and the peg’s tool mount.

Compatibility: the aliases walker, quadruped, and manipulator remain available for older experiment scripts.
WatchRun the live demos →TrainCPU PPO guide →