ToddlerBot - Stanford University's open-source machine learning and humanoid robot platform

What is ToddlerBot?

ToddlerBot is an open-source machine learning and humanoid robot platform from Stanford University designed for efficient collection of large-scale, high-quality training data. It features 30 active degrees of freedom, uses Dynamixel motors, and has a total cost under $6,000. Based on digital twin technology and zero-point calibration, ToddlerBot enables zero-shot transfer from simulation to reality, and its remote operation supports efficient real-world data collection. ToddlerBot excels in motion and manipulation tasks, such as arm span, load, endurance, and dynamic movement capabilities. Its open-source design and detailed assembly manual make it easy to replicate and maintain, suitable for a wide range of research applications.

Main functions of ToddlerBot

• Efficient data collectionIt can collect high-quality training data in both simulated and real-world environments, supporting large-scale machine learning tasks.
• Full-body exercise and gymnasticsIt has 30 active degrees of freedom and can perform complex full-body movements and maneuvers, such as walking, push-ups, pull-ups, bi-arm manipulation, and full-body manipulation.
• Transfer from zero-shot simulation to realityBased on high-fidelity digital twin technology and motor system identification, a seamless strategy transfer from simulation to reality is achieved.
• Remote operation and data collectionEquipped with intuitive remote devices, it supports the rapid collection of real-world data based on human demonstrations for use in learning motor skills.
• Human-computer interaction and collaborationSupports collaborative tasks involving multiple robots, such as working together to clean a room and other complex scenarios.

ToddlerBot's technical principles

• Digital twins and zero-point calibration:
  • Digital twinBased on accurate physical models and system identification technology, high-fidelity simulation models are created to ensure the consistency between simulation data and the real world.
  • Zero point calibrationUsing 3D-printed calibration equipment, the robot's zero-point position can be quickly calibrated to ensure the accuracy of motion control.
• Motor System Identification (SysID)Based on command motor tracking frequency sweep signals, position tracking data is collected, and an execution model is fitted to ensure the accuracy of dynamic parameters. This allows the robot to have the same motion characteristics in both the simulated and real worlds.
• Remote operation technologyUsing a second upper limb as a remote control device, the robot's movement is controlled based on force-sensitive resistors and a handheld gaming PC (such as Steam Deck or ROG Ally X). This allows human operators to intuitively guide the robot in completing complex tasks.
• Reinforcement learning and imitation learning:
  • Reinforcement Learning (RL)Based on the MuJoCo and PPO algorithms, walking and turning strategies are trained, and joint position settings are output to achieve efficient motion control.
  • Imitation learningBased on remote operation, collect real-world data, train a diffusion policy, and implement complex operational tasks.

ToddlerBot's project address

• Project official website:https://toddlerbot.github.io/
• GitHub repository:https://github.com/hshi74/toddlerbot
• arXiv technical paper:https://arxiv.org/pdf/2502.00893

Application scenarios of ToddlerBot

• Home Toy OrganizationTwo robots work together, one picking up toys and the other pushing a cart, to put the toys away.
• Educational programming platformStudents program robots to perform tasks such as walking and pushing up/down.
• Laboratory motor skills researchReinforcement learning trains robots to perform difficult actions such as jumping and climbing.
• Family companionship machinePeople: Interact with children to complete puzzles or sports games.
• Industrial component operation: To operate small electronic components or mechanical parts.