M20 Sim-to-Sim

A lightweight sim-to-sim framework for evaluating trained locomotion policies on the DeepRobotics M20 legged-wheeled robot. Policies are trained in IsaacLab (Isaac Sim) and deployed directly into MuJoCo for cross-simulator validation — no hardware required.

Author: hadelan

---

Overview

m20_sim2sim/
├── deploy/
│   ├── play_mujoco.py              # Main entry point
│   ├── locomotion_policy_wrapper.py # ONNX policy inference wrapper
│   └── utils/mujoco/
│       ├── terrain.py              # Procedural terrain generation
│       └── assets/                 # Scene XML files & heightfield image
├── policy/
│   ├── exported/
│   │   ├── policy.onnx             # Exported policy (ONNX format)
│   │   └── policy.pt               # Exported policy (PyTorch format)
│   └── params/
│       ├── env.yaml                # Training environment config
│       └── agent.yaml              # PPO agent config
└── deep_robotics_model/
    └── M20/
        ├── M20_mjcf/               # MuJoCo MJCF model
        ├── M20_urdf/               # URDF model
        └── M20_usd/                # USD model (for Isaac Sim)

---

Requirements

• Python >= 3.10
• MuJoCo >= 3.0
• mujoco Python bindings
• onnxruntime
• numpy
• pygame (joystick input)
• opencv-python (video recording)
• matplotlib
• pyyaml
• scipy (for Perlin terrain generation)

Install dependencies:

pip install mujoco onnxruntime numpy pygame opencv-python matplotlib pyyaml scipy

---

Quick Start

cd deploy
python play_mujoco.py

Keyboard Shortcuts

Key	Action
A	Activate locomotion policy
S	Deactivate policy (hold standing position)
R	Reset simulation
0	Start / stop video recording
Ctrl-C	Quit and show diagnostic plots

Numpad Directional Control

The right-hand numpad sends persistent velocity commands (press once → stays active until changed).

Key	Command
8	Forward (+vx)
2	Backward (−vx)
4	Strafe left (+vy)
6	Strafe right (−vy)
7	Turn left / CCW (+wz)
9	Turn right / CW (−wz)
5	Stop (clear all numpad commands)

Numpad commands stack additively with joystick input and are reset when the policy is deactivated (S) or the simulation is reset (R).

Joystick Control

If a joystick is detected, use it to command the robot while the policy is active:

Axis	Command
Left stick Y	Forward / backward velocity (±2.0 m/s)
Left stick X	Lateral velocity (±1.0 m/s)
Right stick X	Yaw rate (±1.0 rad/s)

If no joystick is connected, velocity commands are fixed at zero.

---

Configuration

Terrain

Edit the SCENE variable at the top of deploy/play_mujoco.py:

SCENE = 'flat'            # flat ground plane
SCENE = 'random_boxes'    # randomly placed boxes
SCENE = 'random_pyramids' # stacked pyramid obstacles
SCENE = 'perlin'          # Perlin-noise heightfield

Change TERRAIN_SEED for different terrain layouts (same seed = reproducible terrain):

TERRAIN_SEED = 10

Simulation Parameters

In deploy/play_mujoco.py:

Parameter	Default	Description
SIM_DT	5e-3 s	Physics timestep
DECIMATION	4	Policy runs every N sim steps (policy freq = 50 Hz)
RENDER_HZ	30	Viewer render frequency
RAMP_STEPS	200	Smooth ramp-up steps when activating policy

Policy & Actuator Parameters

Defined in deploy/locomotion_policy_wrapper.py:

Parameter	Value	Description
KP_LEG	80.0	Leg joint position gain
KD_LEG	2.0	Leg joint velocity gain
KD_WHEEL	0.6	Wheel velocity gain
EFFORT_LEG	76.4 Nm	Leg joint torque limit
EFFORT_WHEEL	21.6 Nm	Wheel torque limit

Swapping the Policy

Replace the files in policy/exported/ with your own exported policy:

• policy.onnx — required for MuJoCo inference
• policy.pt — optional PyTorch checkpoint

The observation space expected by the wrapper (57 dims):

[0:3]   base angular velocity (body frame) × 0.25
[3:6]   projected gravity (body frame) × 1.0
[6:9]   velocity commands [vx, vy, wz] × 1.0
[9:25]  joint positions relative to default (wheels zeroed) × 1.0
[25:41] joint velocities × 0.05
[41:57] last action × 1.0

---

Diagnostic Plots

After quitting (Ctrl-C), the following plots are displayed automatically:

• Leg joint torques over time (per joint, per leg)
• Leg joint velocities over time
• Torque–velocity operating point scatter (colour = power)
• |Torque|–|velocity| scatter with constant-power curves
• Wheel velocity and torque over time

---

Video Recording

Press 0 in the viewer to start/stop recording. Videos are saved as m20_video_<timestamp>.mp4 in the working directory.