本文由 AI 分析生成
建立時間： 2026-03-28 來源： https://toddlerbot.github.io
Summary
Shi, Wang, Song, and Liu (Stanford, CoRL 2025) present ToddlerBot, a low-cost open-source miniature humanoid platform specifically designed for ML-compatible policy learning research. At 0.56m/3.4kg with 30 active DoFs and a cost under $6,000, it fills the gap between expensive commercial humanoids and limited research platforms — while providing full sim-to-real infrastructure including digital twin, sysID pipeline, and teleoperation data collection.
Shi、Wang、Song 和 Liu（史丹佛，CoRL 2025）提出 ToddlerBot，一個低成本開源小型人形機器人平台，專為 ML 兼容的策略學習研究設計。在 0.56m/3.4kg、30 個主動自由度、成本不超過 6,000 美元的條件下，填補了昂貴商業人形機器人和有限研究平台之間的空白——同時提供完整的仿真到現實基礎設施，包括數字孿生、系統辨識流程和遙操作數據收集。
Prerequisites
- Reinforcement learning and sim-to-real transfer
- Robot kinematics and dynamics modeling
- System identification (sysID) methods
- Teleoperation and visuomotor policy learning
Core Idea
Existing humanoid platforms fall into two categories: industrial-scale robots (300K, closed-source, impractical for daily research iteration) or minimal research platforms (few DoFs, limited manipulation capability). ToddlerBot bridges this gap with three design pillars:
- ML-compatibility: plug-and-play zero-point calibration + transferable motor sysID ensures high-fidelity digital twin; supports zero-shot sim-to-real policy transfer; intuitive teleoperation interface for real-world data collection
- Capability: 30 active DoFs (more than any comparable miniature humanoid); anthropomorphic design enables transfer of human demonstrations; validated on walking, push-ups, pull-ups, wagon pushing, bimanual and full-body manipulation
- Reproducibility: entirely 3D-printed, off-the-shelf components ($6K total, 90% motors and computers); validated by independent replication by a CS student not on the project team; 5 independent replication reports within one week of release
Results
- Demonstrated loco-manipulation skills: walking (RL policy, zero-shot sim-to-real), push-ups and pull-ups (keyframe interpolation), wagon pushing, bimanual manipulation, full-body manipulation
- Collaborative scenario: two robots tidying a toy session together
- Superior DoF count vs. comparable price point: 30 DoFs at 5K), Zeroth (16 DoFs at 6.5K)
| Platform | Height | Weight | DoFs | Price | Open Source |
|---|---|---|---|---|---|
| ToddlerBot (ours) | 0.56m | 3.4kg | 30 | $6K | design+code |
| Berkeley Humanoid Lite | 0.80m | 16kg | 22 | $5K | design+code |
| Zeroth | 0.48m | 3.6kg | 16 | $1.4K | design+code |
| NAO H25 | 0.57m | 5.2kg | 23 | $14K | code only |
Limitations
Author-stated:
- Compact size limits payload capacity compared to full-sized humanoids
- Cost breakdown (90% motors/compute) means the BOM is sensitive to motor prices
Unstated:
- 3D-printed structure may have durability limitations under repeated high-force manipulation
- 3.4kg is extremely light — may limit applicability to tasks requiring stable base under load
- Digital twin fidelity depends on sysID accuracy; real-world deployment may still require fine-tuning
Reproducibility
- Code: Available at toddlerbot.github.io
- Hardware: Fully open-source design (3D-printable, off-the-shelf BOM)
- Venue: CoRL 2025, Seoul, Korea; arXiv: 2502.00893
Insights
ToddlerBot’s most important contribution may be the sysID pipeline and digital twin infrastructure rather than the physical design itself. Zero-shot sim-to-real transfer at a miniature scale is non-trivial; the calibration and identification methodology that makes it work is what enables researchers to scale data collection through simulation. The fact that five independent teams replicated the system within a week of release validates the reproducibility claim in a way few robot papers achieve. This platform directly addresses the “data problem” for humanoid research by being cheap enough to build multiple copies and instrumented enough to collect high-quality data at scale.
Connections
Raw Excerpt
ToddlerBot, a low-cost, open-source humanoid robot platform designed for robotics and AI research. ToddlerBot enables seamless acquisition of high-quality simulation and real-world data. The plug-and-play zero-point calibration and transferable motor system identification ensure a high-fidelity digital twin and zero-shot sim-to-real policy transfer.