humanoidrobots.systems

stanford/toddlerbot · spec sheet

Stanford ToddlerBot specs

Low-cost, fully 3D-printed open-source small humanoid (0.56 m, 3.4 kg, 30 DOF) for scalable policy learning research.

Maker
Class
Humanoid
Status
Research
Availability
DIY build
Country
United States, Stanford
Level
L2 Sim-verified
01

Media

4 items · credited and licensed
02

Specifications

7/12 fields sourced
Stanford ToddlerBot specifications
Heightspecs.heightCm56cm
Massspecs.massKg3.4kg
Degrees of freedomspecs.dof30
Payloadspecs.payloadKg1.48kg
Top speedspecs.speedMpsunsourced
Runtimespecs.runtimeMinunsourced
Battery capacityspecs.batteryWhunsourced
Pricespecs.priceUsdunsourcedBuild cost under $6,000 (paper); not sold
Actuatorsspecs.actuatorsservo actuators; 7 DOF/arm, 6 DOF/leg, 2 neck, 2 waist
Computespecs.computeNVIDIA Jetson Orin NX 16GB
Sensorsspecs.sensorsunsourced
Year introducedyear2025
Licenselicense.hw / .sw / .docsHardware: CC-BY-NC-SA-4.0Software: MIT

Note · ToddlerBot 2.0 released Aug 2025; variants 2XC and 2XM both in Menagerie (toddlerbot_2xm). Onshape CAD + MakerWorld print files + BOM sheet public. Locomotion Beyond Feet follow-up Jan 2026. Mirror fork exists at kscalelabs/toddlerbot. Licenses per README: code and docs MIT; hardware design (Onshape, STL/3MF) CC BY-NC-SA 4.0, non-commercial.

03

Openness

method
100 / 100Open hardware
ArtifactPublicWeightPoints
CADopenness.cadyes2525
Bill of materialsopenness.bomyes2020
Sim model (URDF/MJCF)openness.simyes2020
Firmwareopenness.firmwareyes1515
Softwareopenness.softwareyes1010
Trained policiesopenness.policiesyes1010
score = Σ weight × public100
04

Build package

v0.1.0 · full guide
BOM lines
47
Parts (qty)
342
Priced BOM
$5,728
Fab parts
13
Print time
54.6 h
Material
1.5 kg
Build steps
16
Est. assembly
—
Safety steps
6
Vendors
11

BOM by category · USD from priced lines

CategoryLinesQtyUSD
Actuators5223,908
Compute11955
Structure648241
Power56171
Sensors23125
Fasteners920994
Printed2390
Cabling94273
Electronics6649
Other2224

Offers in CN · GLOBAL · US from 11 vendors

Subsystems · assembly groups

GroupStepsMinSafety
Fabrication3——
Arms1——
Legs1——
Head1——
Waist1——
Torso1——
Bring-up4—3
Compute1——
Electrical2—2
Software1—1

Fabrication: FDM print ×13

Firmware

Targets
Dynamixel X-series servos (Protocol 2.0, TTL, 2 Mbps), TTL power board V8 (CH348L 8-port USB-to-TTL, no MCU firmware), Jetson Orin NX host (JetPack 6.1+, RT kernel)
Repo
github.com/hshi74/toddlerbot
Calibrate
python toddlerbot/tools/calibrate_zero.py --robot toddlerbot_2xc [--parts ...]; python toddlerbot/descriptions/add_configs.py --robot toddlerbot_2xc; fine-tune: python toddlerbot/policies/run_policy.py --policy calibrate --robot toddlerbot_2xc --sim real
Notes
No custom motor firmware. Motor IDs and baud rate are set by hand in ROBOTIS Dynamixel Wizard 2.0 (manual + 'ID the Motors' video); the repo has no ID script. Host control is a C++ pybind11 extension (toddlerbot/actuation/src/dynamixel_mch.cpp + vendored dynamixel_sdk) that scans the serial ports for motor IDs; toddlerbot/sim/real_world.py opens them at 2,000,000 baud. examples/test_dynamixel.py tests the bus. Per-motor system ID: toddlerbot/tools/run_sysID.py with the SysID jig (MakerWorld plate 7); cached fits are in toddlerbot/descriptions/sysID_*/. Per-robot configs: toddlerbot/descriptions/toddlerbot_2xc/ (MJCF, URDF, robot.yml).

Policies

  • walk (omnidirectional, 2XC) · MuJoCo MJX + RSL-RL PPO, ONNX Runtime on Jetson · drive.google.comCheckpoint toddlerbot_2xc_walk_rsl_20251226_114612.zip in the Google Drive folder linked from the README. It is also the WALK_POLICY used by run_multiple_policy.py.
  • get up from the ground · MuJoCo MJX + RSL-RL PPO · drive.google.comCheckpoint toddlerbot_2xc_get_up_rsl_20260214_145222.zip (Google Drive).
  • crawl (full / only / rotate) · MuJoCo MJX + RSL-RL PPO · drive.google.comCheckpoints toddlerbot_2xc_crawl_full_rsl_20260214_145000.zip, crawl_only_rsl_20260214_145735.zip, crawl_rotate_rsl_20251109_195839.zip (Google Drive). Entry scripts crawl_full.py, crawl_only.py, crawl_rotate.py.
  • crawl up / down stairs · MuJoCo MJX + RSL-RL PPO · drive.google.comCheckpoints toddlerbot_2xc_crawl_up_stairs_rsl_20260217_211836.zip and crawl_down_stairs_rsl_20260215_094639.zip (Google Drive). Down-stairs entry: crawl_down_stairs.py.
  • climb up / down box, climb wall · MuJoCo MJX + RSL-RL PPO · drive.google.comCheckpoints toddlerbot_2xc_climb_up_box_rsl_20250907_201202.zip, climb_down_box_rsl_20250914_014343.zip, climb_wall_rsl_20250913_174540.zip (Google Drive). Entry scripts climb_down_box.py, climb_wall.py.
  • multi-skill whole-body locomotion (Locomotion Beyond Feet) · RL skill policies + depth skill classifier (FoundationStereo) · github.comChains the Drive checkpoints through a depth-image skill classifier. No classifier checkpoint is published; train it with toddlerbot/skill_classifier/training/train.py on your own collected data.
  • push-up (keyframe replay) · open-loop keyframe replay (run_policy.py --policy replay) · github.commotion/ holds 2XC and 2XM keyframe/reference motions (cartwheel, crawl, get-up, pull-up, kneel, walk_zmp, ...). These are reference motions, not trained networks.
  • cartwheel (DeepMimic-style) · MuJoCo MJX + RSL-RL PPO · github.comNo cartwheel checkpoint in the Drive folder. Train with train_mjx.py --gin-file cartwheel --env cartwheel against motion/cartwheel_2xc.lz4. The changelog says the success rate is still low.
  • bimanual manipulation (diffusion policy) · Diffusion Policy (PyTorch) · github.comExpects results/<robot>_<task>_dp_<ckpt>/best_ckpt.pth from your own teleop data; no DP weights are published.

Caveats · stated by the package

  • Variant: this package is toddlerbot_2xc, the docs' 'recommended default' and the only variant with published RL checkpoints. toddlerbot_2xm (stronger hips: 6x 2XC430, 0x 2XL430, +4x XM430-W350; separate MakerWorld release) is noted in the alts; the sheet's 2XM essentials total is $7,293.56. Sources disagree on which to build: MakerWorld's 2XC page calls 2XM 'recommended for all new builds too'.
  • Licensing: code and docs are MIT, but the hardware design (Onshape, STL/3MF) is CC BY-NC-SA 4.0 (non-commercial) per the README and MakerWorld.
  • The assembly manual PDF in the repo is the 1.0 manual, and the docs say so: 2.0 users are sent to overview videos, Onshape and a tnkr.ai interactive manual. The tnkr.ai link (docs.tnkr.ai/Stanford University) did not respond on 2026-09-25. Steps here follow the 1.0 manual and 2.0 videos; motor counts per subassembly are from the 1.0 manual.
  • The 1.0 manual assigns IDs with a U2D2 + power hub + 12 V adapter. The 2.0 essentials list neither (only under Teleoperation, $32.10 + $19.00), and no 12 V adapter is listed anywhere.
  • The per-joint ID list and chain diagram exist only as a figure in the manual PDF. The repo has no ID-assignment script and no plain-text ID table.
  • Bearings: the Screws tab says 683 x12 are used, but the sheet buys one 10-pack. Buy two.
  • Plate 16 'Palm with TPU' uses 62 g of TPU, but the BOM lists no TPU. This package uses the rubber-tape palm (plate 18) instead.
  • MakerWorld plates carry no per-part lists, so fabricated qty is per plate, not per part. Plate 19 is unlabelled. Optional plates are excluded: Safety Stand, SysID, Teleop, Gripper.
  • Printed parts do not appear in the repo, which only has visual/collision meshes under descriptions/. Printable files live on MakerWorld (Bambu 3MF) and Onshape.
  • Prices are the undated USD figures in the published sheet. ROBOTIS prices disagree inside the sheet (2XL430 $149.39 in Essentials vs $129.90 under Teleoperation), so treat them as stale. JST EH and M2x6 unit prices are derived from the sheet's line totals ($3.82, $1.17, $1.80). The priced lines add up to the sheet's 2XC essentials total ($5,727.94), which excludes the PCB, gripper, tools and optional features.
  • The 8-channel TTL power board has no upstream price. The docs say the JLCPCB cost is mostly setup fees and varies with tariffs; order 5 (2 minimum).
  • Several vendor links are short or share links that could not be expanded or verified: DigiKey /short/ links returned 403, and the Taobao e.tb.cn link. The LiPo listing's cell count and voltage are not stated in the sheet.
  • Tools are not in this BOM. The sheet lists about $1,976 of tools (Bambu X1C, bench supply, screwdrivers, soldering kit) and about $2,253 of optional features: safety stand, arm teleop leader, DIY 21700 battery, SysID clamp, ROG Ally X.
  • An assembled 2.0 is sold by WowRobo ($4,299 at shop.wowrobo.com, pre-order, sold out on 2026-09-25; Jetson, Wi-Fi and battery not included). The sheet disclaims any affiliation with it.
  • The published checkpoints are Google Drive zips from the authors (all 2XC), not versioned in the repo. Some loaders (mjx_policy.load_wandb_policy) default to a private W&B entity, 'toddlerbot'.

Package stanford/toddlerbot@0.1.0 · upstream hshi74/toddlerbot@e337f3b177b4b53abff70b31d1695a7b66cc6d2e (2026-04-19; ToddlerBot 2.0.0 released 2025-08-25). BOM Google Sheet "Release" + "Screws" tabs read 2026-09-25 (undated sheet). MakerWorld 2XC print profile last noted update 2025-11-21. Package targets toddlerbot_2xc. · 25 package sources

05

Models, repositories, links

6 links
06

Videos

2 tagged · 0 from the maker's channel
ToddlerBot: Open-Source ML-Compatible Humanoid Platform for Loco-ManipulationHaochen ShiOfficial embedyoutube.comStanford UniversityStanford ToddlerBot
ToddlerBot 2.0 Arm OverviewHaochen ShiOfficial embedyoutube.comStanford UniversityStanford ToddlerBot
07

News and papers

0 stories · 0 papers

News mentioning ToddlerBot

No story in the news feed is tagged ToddlerBot yet.

Papers using ToddlerBot

No paper in the arXiv feed is tagged ToddlerBot yet.

20251

  1. Open source

    Stanford releases ToddlerBot

    Stanford researchers described ToddlerBot, a low-cost open-source humanoid designed for machine-learning research in locomotion and manipulation.

    Stanford UniversityStanford ToddlerBotarxiv.orgtoddlerbot.github.io
09

Sources

last checked 2026-09-25
  1. [1]Official repo (README license section: code MIT, design CC BY-NC-SA 4.0)github.com/hshi74/toddlerbotchecked 2026-09-25
  2. [2]arXiv paperarxiv.org/abs/2502.00893checked 2026-09-24
  3. [3]Project sitetoddlerbot.github.io/checked 2026-09-24
10

Suggest a correction

Corrections need a public source. They go to the Hyperspawn intake queue and are reviewed before the record changes; the page updates on the next data release.

now: 56
POST hyperspawn.io/api/v1/reports
11

Related

12

Cite this page

humanoidrobots.systems. “Stanford ToddlerBot specs.” Hyperspawn, data revision 2026-11-13. https://humanoidrobots.systems/stanford/toddlerbot

Machine-readable: /data.json (record stanford/toddlerbot).