Retarget Dataset
Chicken Dance
Jump Motion
We've brought more natural and interactive movements to the Adamlite robot by retargeting motion data from the SFU dataset, along with some of our own captured and created motions. This allows the robot to perform the natural walking and interactive behaviors you see in our public videos.
This was all made possible using a whole-body inverse kinematics (IK) numerical optimization approach. Our optimization focused on joint position and joint velocity constraints to ensure the movements are both realistic and stable. For some shorter actions, we've also repeated the motion multiple times to extend their usability.
Requirements & System Specifications
To ensure the visualization and simulation run correctly, your system should meet the following requirements:
System Requirements
- OS: Linux Ubuntu 18.04 / 20.04 / 22.04 (Isaac Gym does not support Windows/macOS natively).
- GPU: NVIDIA GeForce RTX 20 series or higher (RTX 3060+ recommended).
- Driver: NVIDIA Driver version >= 470.xx.
Software Environment
- Python: 3.8 (Required for Isaac Gym Preview 4 compatibility).
- Conda: For environment management.
- CUDA: 11.7 (Aligned with the PyTorch version used).
- Isaac Gym: Preview 4 version.
Trajectory Data Format
Our released trajectories come in a comprehensive JSON format. Each file provides:
- Frame rate: This tells you the time resolution of the motion.
- Data labels: These clearly identify the different data streams within the trajectory.
- Trajectory data: This is the core motion information, stored as a 2D array.
Within this trajectory data, you'll find detailed insights into the robot's state and movement, including:
- Base velocity: The robot's linear and angular speed at its base.
- Base pose: The robot's position and orientation in space.
- Joint angles: The angular positions of each joint.
- Joint velocities: The angular speeds of each joint.
Important Note on Execution
It's crucial to understand that these trajectories were developed with only kinematic constraints in mind, meaning we focused on the robot's physical structure and range of motion. Dynamic considerations (like forces and torques) were not included. Because of this, these trajectories cannot be directly executed perfectly on the robot without further dynamic control and adjustments.
visualize robot trajectories
# Step 1: Create a new python virtual env with python 3.8
conda create -n retarget python=3.8
conda activate retarget
# Step 2: Install pytorch 1.13 with cuda-11.7:
conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia
conda install numpy=1.23
# Step 3: Clone repository and Install IsaacGym
# 3.1 git clone
cd IsaacGym_Preview_4_Package/isaacgym/python && pip install -e .
# Step 4: Run the script
# run the script with parameters:
python visualize.py --robot_type=adam_lite --file_name mixamo/low_jump.json
python visualize.py --robot_type=adam_sp --file_name lafan1/walk4_subject1_extended.csv
Dataset Collection Pipeline
The Order of Configuration
Adam lite(23 DOF channels):
- root_pos/x
- root_pos/y
- root_pos/z
- root_quat/x
- root_quat/y
- root_quat/z
- root_quat/w
- dof_pos/hipPitch_Left
- dof_pos/hipRoll_Left
- dof_pos/hipYaw_Left
- dof_pos/kneePitch_Left
- dof_pos/anklePitch_Left
- dof_pos/ankleRoll_Left
- dof_pos/hipPitch_Right
- dof_pos/hipRoll_Right
- dof_pos/hipYaw_Right
- dof_pos/kneePitch_Right
- dof_pos/anklePitch_Right
- dof_pos/ankleRoll_Right
- dof_pos/waistRoll
- dof_pos/waistPitch
- dof_pos/waistYaw
- dof_pos/shoulderPitch_Left
- dof_pos/shoulderRoll_Left
- dof_pos/shoulderYaw_Left
- dof_pos/elbow_Left
- dof_pos/shoulderPitch_Right
- dof_pos/shoulderRoll_Right
- dof_pos/shoulderYaw_Right
- dof_pos/elbow_Right
- dof_vel/hipPitch_Left
- dof_vel/hipRoll_Left
- dof_vel/hipYaw_Left
- dof_vel/kneePitch_Left
- dof_vel/anklePitch_Left
- dof_vel/ankleRoll_Left
- dof_vel/hipPitch_Right
- dof_vel/hipRoll_Right
- dof_vel/hipYaw_Right
- dof_vel/kneePitch_Right
- dof_vel/anklePitch_Right
- dof_vel/ankleRoll_Right
- dof_vel/waistRoll
- dof_vel/waistPitch
- dof_vel/waistYaw
- dof_vel/shoulderPitch_Left
- dof_vel/shoulderRoll_Left
- dof_vel/shoulderYaw_Left
- dof_vel/elbow_Left
- dof_vel/shoulderPitch_Right
About PND Adam
- The whole body is composed of up to 41 quasi-direct drive flexible force-controlled actuators.
- Adam Lite: 25 degrees of freedom.
- Adam Standard: 29 degrees of freedom.
- Adam SP: 41 degrees of freedom.
- Height: 1.67m, weight: 60kg.
- The legs use four quasi-direct drive force-controlled joints with the highest torque density in the industry and a 7-speed ratio and high sensitivity.
- The arms have up to 19 degrees of freedom.
- Adam Lite: 5 degrees of freedom.
- Adam Standard: 7 degrees of freedom.
- Adam SP: 19 degrees of freedom.
- The waist has 3 degrees of freedom.
Introduction to the Appearance of Humanoid Robots
Names of Main Components
