v2.0 Sim-to-Real Toolkit

Open Locomotion
Skills Dataset

A unified multi-embodiment foundation and sim-to-real toolkit for legged robots. Featuring hardware ingestors for Unitree and ANYmal, procedural terrains, and a strict Go1 baseline benchmark.

View Benchmark Quick Start
0.61 m/s
Forward Velocity
12m
Flat Displacement
99.9%
Sim-to-Real Alignment
3
Procedural Terrains
v2.0
Release

Go1 Velocity-Tracking Benchmark

Genuine locomotion at 0.6 m/s. Per-episode velocity randomization for robust generalization across terrains.

Baseline▲ Terrain▲ Success Rate▲ Fall Count▲ Ep. Length▲ Fwd Velocity▲
OLSD v2 veltrack
Flat
1.0
 0 1000.0
0.6108 m/s
OLSD v2 veltrack
Stairs
0.8
 4 874.2 0.0893 m/s
OLSD v2 veltrack
Slope (in progress)
0.0
 — 228.8 —

Supported Real & Sim Robots

Expanding beyond simple planar agents into real-world quadruped and humanoid hardware architectures.

ANYmal-D
Quadruped / 12 DOF
Real
Data Source
12
Joints
99.9%
Alignment
Unitree Go1
Quadruped / 12 DOF
Sim
Data Source
12
Joints
PPO
Baseline
Unitree G1
Humanoid / 23 DOF
Mocap
Data Source
23
Joints
TBD
Baseline
TAIL
Quadruped / 12 DOF
Mixed
Data Source
12
Joints
Sand
Terrain

Quick Start

Generate, train, and evaluate in a few commands.

terminal
# Install pip install -e ".[sim,dev]"
# Train a Go1 model with sim-to-real configs python scripts/train_go1.py \ --terrain stairs \ --output-dir checkpoints/go1-stairs \ --sim2real-config configs/sim2real/go1.yaml # Compare head-to-head strict baselines python scripts/compare_go1_baselines.py \ --flat-policy checkpoints/phase2/go1/flat/selected_model.zip \ --sim2real-config configs/sim2real/go1.yaml \ --output results/go1_head_to_head.json
python
import olsd from olsd.generation.mujoco_gen import generate_trajectories # Generate 50 expert episodes episodes = generate_trajectories("halfcheetah", n_episodes=50) # Load the canonical dataset dataset = olsd.load("./data/olsd-v0.1-final") print(dataset.summary())