Python-first scenes
Create geometry, articulated actors, sensors, actuators, terrain, and task logic through a validated Python DSL.
Native physics · Python control · Web visualization
Robotics worlds,
authored in Python.
Define bodies, joints, sensors, rewards, and interaction loops in regular Python. Run the same compact C++ physics core on macOS, Linux, and WebAssembly.
$ pip install lavendersim from lavendersim.env import make env = make("RoverWaypoint-v0", seed=7) obs, info = env.reset(seed=7) obs, reward, terminated, truncated, info = \ env.step(env.action_space.sample()) print(obs.shape, reward) # (28,) 0.294...
Chapter one: make a tiny world, then teach it something curious.
13registered RL tasks
32×default CPU rollout workers
3native targets: Linux, macOS, WASM
0XML required
Actual native camera output
Small scenes. Real control problems.
These are not concept art. They are deterministic RGB frames rendered by LavenderSim’s C++ camera sensor from registered Python environments.
A small, practical stack
From scene definition to trained policy.
LavenderSim keeps task logic visible and hackable while moving physics into a compact native engine.
Robotics observations
IMUs, tactile grids, range fans, cameras, named sites, contact force/torque, noise, filtering, and delay.
Useful actuation
Position, velocity, effort, impedance, gear and force limits, activation dynamics, and fixed tendon coupling.
CPU PPO included
Flat policy observations, seeded randomization, process-vectorized rollouts, checkpoints, and evaluation tools.
Deterministic snapshots
Save simulation, sensor, actuator, time, and RNG state. Restore it to compare alternate action sequences.
Python-owned web UI
Stream authoritative native poses to the browser and send motion commands or clicked targets back to Python.
The full loop
Policy training and visualization share one source of truth.
Python defines what the task means. The native engine advances physics and sensors. The browser displays authoritative poses instead of running a second, drifting simulation.
Benchmarks included
Learn balance, locomotion, sensing, and manipulation.
Quadruped terrain
Track randomized body-relative direction and speed using 12 joints, an IMU, and terrain probes.
Peg insertion
Align named frames while using filtered contact force and torque to control insertion.
Rover waypoint
Drive over a heightfield toward randomized goals with differential wheels and ray-fan lidar.
Experimental and AI-generated. Almost the entire project was generated by GPT-5.6 Sol under human direction. LavenderSim is alpha research software—not a validated MuJoCo replacement or a safety-certified robotics system.