AI Wiki

Explore AI Wiki

Citation and evidence

SO-101

16 min full readUpdated 25 references

This article's verification

Report a problem with this article

More

Use this article

Raw MarkdownExplore connections

Improve this page

Suggest editRevision historyDiscussion

Browse categories

Embodied AIOpen Source AIRobot HardwareRobotics

Cite this article

The SO-101 (Standard Open Arm 101) is a low-cost, open-source, 3D-printable robot arm designed by The Robot Studio in collaboration with Hugging Face and published in the TheRobotStudio/SO-ARM100 repository under the Apache 2.0 license.[1][2] It is the successor to the SO-100, the arm the same partners released in 2024, and it was announced on 28 April 2025 alongside kit vendors WowRobo, Seeed Studio and PartaBot.[3][4] The arm is built from printed plastic parts around six Feetech STS3215 serial-bus servos and a single USB servo-driver board, and it is normally used as a pair: a hand-held "leader" arm whose joint positions are read out, and a motorised "follower" arm that mirrors them.[5] That pairing is the standard teleoperation rig for recording demonstrations with Hugging Face's LeRobot library, which treats the SO-101 as its flagship robot.[5]

The project's own bill of materials priced a single follower arm at $121.94 and a leader-plus-follower pair at $229.88 in the United States (parts only, excluding the 3D printer), as listed in the README on 16 September 2026; the same list quotes 124.30 EUR and 226.30 EUR in Europe.[1] LeRobot's 2026 paper gives roughly 225 EUR for one SO-100/101 teleoperation setup and about 550 EUR for a bimanual one, against roughly 21,000 EUR for an ALOHA bimanual system.[6] On the launch day, Hugging Face co-founder Clement Delangue said kits "cost from $100 to $500 depending on how much you want it assembled and your country of shipping," and LeRobot's Remi Cadene quoted "$100 per arm" and a 20-minute assembly.[7][8] The GitHub repository had 7,475 stars and 685 forks on 16 September 2026.[2]

Origins and naming

The Robot Studio, a France-registered studio run by roboticist Rob Knight, lists him first in the project's citation file; the other ten authors (Pepijn Kooijmans, Remi Cadene, Simon Alibert, Michel Aractingi, Dana Aubakirova, Adil Zouitine, Russi Martino, Steven Palma, Caroline Pascal and Thomas Wolf) all appear as Hugging Face authors on the LeRobot or SmolVLA papers.[6][9][16] The README describes the SO-100 as "the original robot arm designed by the RobotStudio in collaboration with Hugging Face," and the SO-101 as its "next-generation version."[1][10]

The repository name, SO-ARM100, predates the SO-101 and is still used for both generations; the CAD files for the first design carry the label SO_5DOF_ARM100, and TechCrunch, Hugging Face and the LeRobot documentation all refer to the arms as SO-100 and SO-101.[1][11] LeRobot's 2026 paper refers to the two jointly as "SO-10X."[6] This article uses SO-101 for the current design and SO-100 for the original.

History

DateEventSource
10 May 2024TheRobotStudio/SO-ARM100 repository created ("Initial commit")[2]
16 May 2024Changelog v0.1.0: first documented revisions of the printed jaw and wrist parts[11]
25 Sep to 1 Oct 2024Changelog v0.1.6 to v0.1.11: URDF for the SO_5DOF_ARM100_8j design, print trays sized for Ender and Prusa beds, SolidWorks CAD, and STL files updated to the SO_5DOF_ARM100_08k design[11]
1 Oct 2024Changelog v0.1.12: "URDF for early draft of 7DOF SO-Arm to allow modelling to begin"[11]
25 Oct 2024LeRobot merges "Add FeetechMotorsBus, SO-100, Moss-v1" (#419), adding SO-100 support to the library[12]
4 Feb 2025LeKiwi mobile-manipulator repository created at SIGRobotics-UIUC, using an SO-ARM101 arm[13]
28 Apr 2025Pull request #67 "Add SO101, cleanup repo" adds SO-101 STEP and STL files, moves the SO-100 documentation to SO100.md, and Hugging Face announces the SO-101[3][7][14]
12 May 2025SO-101 MuJoCo (MJCF) and URDF files added (#80)[15]
22 May 2025CITATION.cff added (#90)[9]
2 Jun 2025SmolVLA paper reports real-world results on SO-100 and SO-101[16]
16 Jul 2025LeRobot adds a bimanual SO-100 robot configuration (#1509)[17]
26 Feb 2026LeRobot library paper reports SO-100 and SO-101 as the two robots with by far the most openly shared datasets on the Hugging Face Hub[6]

Expanded article table

The SO-100 documentation is retained in SO100.md but is marked "deprecated" in the current README.[1][10]

What changed from SO-100 to SO-101

The README lists three changes: improved wiring, easier assembly ("no gear removal"), and updated motors for the leader arm.[1] The SO-100 build used twelve identical STS3215 servos for both arms; the SO-101 keeps six 1/345-geared servos for the follower but gives the leader arm three different gear ratios so that it "can both sustain its own weight and it can be moved without requiring much force."[5][10] The LeRobot SO-100 page notes a practical consequence of the redesign: on the SO-100 "the motor connectors are not easily accessible once the arm is assembled, so the configuration step must be done beforehand," whereas the SO-101 assembly guide sets each motor's ID one motor at a time before the parts are assembled, with the connectors staying accessible afterwards.[4][5] The SmolVLA paper summarises the difference as "better arm design for faster assembly and different motors, making its movements smoother and better for tasks requiring more precision."[16] Cadene's launch post claimed assembly in 20 minutes.[8]

Specifications

The project publishes a bill of materials, print settings, CAD and simulation files, but no datasheet-style figures for reach, rated payload, joint speed or repeatability. Where a value is not published by the project, the table says so rather than substituting a reseller's number.

ItemSO-101 followerSO-101 leaderSource
Actuated joints6: shoulder pan, shoulder lift, elbow flex, wrist flex, wrist roll, gripper (parallel jaw)6, with a trigger handle in place of the gripper[5][18]
Degrees of freedomDescribed as 6-DOF by NVIDIA and the SmolVLA paper; the project's own CAD files are labelled SO_5DOF_ARM100 (five arm joints plus gripper)same[11][16][18]
Servos6 x Feetech STS3215, 7.4 V, 1/345 gear (part C001)1 x 1/345 (C001, shoulder lift); 2 x 1/191 (C044, shoulder pan and elbow flex); 3 x 1/147 (C046, wrist flex, wrist roll, gripper)[1][5]
Servo torque7.4 V STS3215: 16.5 kg.cm stall torque at 6 V; optional 12 V version: 30 kg.cm7.4 V only[1]
Controller1 x "Motor Control Board" serial-bus servo driver (the US BOM links a Waveshare board), USB-C to hostsame[1]
Power5 V supply per arm; 12 V 5 A or more if 12 V servos are used5 V[1]
Structure3D-printed PLA+, 0.4 mm nozzle at 0.2 mm layers or 0.6 mm nozzle at 0.4 mm layers, 15 percent infill; 9 shared parts plus 2 follower-specific parts9 shared parts plus 3 leader-specific parts (handle, trigger, wrist roll)[1]
Print bedSingle-file trays for 220 x 220 mm (Ender) and 205 x 250 mm (Prusa/UP) bedssame[1]
Reach, payload, repeatability, speedNot published by the projectNot published[1]
CAD and simulationSTEP and STL files; URDF (old and new calibration) and MuJoCo MJCFSTEP and STL[15]
LicenseApache 2.0Apache 2.0[2]
Parts cost (US, project BOM, 16 Sep 2026)$121.94 for one follower$229.88 for the leader-plus-follower pair[1]

Expanded article table

Optional add-ons documented in the repository include wrist-camera mounts for 32 x 32 mm UVC modules, Intel RealSense D405 and D435/D435i, and a Vinmooog webcam; an overhead camera mount for single and bimanual setups; a raised leader base and a 4040 aluminium-profile base mount; a printed mount-alignment jig; compliant TPU gripper fingers; and a listed AnySkin tactile sensor from WowRobo.[1]

Leader-follower teleoperation and data collection

In the LeRobot workflow each arm is first assigned a USB port (lerobot-find-port), then each servo is given a unique bus ID and baud rate one at a time (lerobot-setup-motors), and finally both arms are calibrated by moving every joint through its range so that leader and follower report the same values in the same physical pose.[5] The docs stress that calibration "allows a neural network trained on one robot to work on another."[5] Once calibrated, the follower is driven by so101_follower and the leader by so101_leader robot and teleoperator types; the operator moves the leader by hand and the follower copies it while cameras, joint states and actions are written into the LeRobotDataset format, which can be pushed to the Hugging Face Hub.[5][6]

That loop is what made the SO-10X arms the main source of community robot data on the Hub. In the LeRobot paper's Hub statistics, SO-101 datasets had been downloaded 319,586 times across 3,965 datasets and 58,299 episodes, and SO-100 datasets 278,697 times across 5,161 datasets and 78,510 episodes; in the paper's main comparison table only the Franka Panda, xArm, WidowX and KUKA embodiments (all backed by large academic datasets) had more downloads, although its appendix also lists the Google robot and an "unknown" embodiment group above the SO arms; no other robot came close in number of datasets.[6] The paper also notes that over 2025 LeRobot grew from three supported manipulation setups (Koch v1.1, SO-100, ALOHA) to eight, and that it supports the SO-100 and SO-101 "both in a single and bimanual setup."[6]

Policies documented as trained or evaluated on SO-100/SO-101 hardware include:

PolicyEvidenceSource
ACTBaseline in the SmolVLA real-world tables; 19/20 successes in a bimanual SO-101 beanbag task on a Jetson Orin Nano Super; baseline in the SO-101 VLA benchmark[16][22][20]
SmolVLAPretrained on 481 community datasets (22.9K episodes, 10.6M frames) collected on SO-100 arms; evaluated on SO-100 and SO-101[16]
π0Multi-task baseline on SO-100 in the SmolVLA paper[16]
π0.5 and Wall-XFine-tuned and evaluated on SO-101 in the VLA failure-and-recovery benchmark[20]
Diffusion policyTrained on the same bimanual SO-101 demonstrations as ACT in the Jetson study, where it did not converge at its 200k-step budget[22]
NVIDIA Isaac GR00TUsed in NVIDIA's "Train an SO-101 Robot From Sim-to-Real With NVIDIA Isaac" learning path[18]

Expanded article table

SmolVLA results on SO-100 and SO-101

SmolVLA is the clearest example of the arm feeding a model. Its pretraining corpus was 481 Hub datasets filtered by embodiment, episode count and quality, all recorded on SO-100 arms; the authors list single-embodiment pretraining as a limitation.[16] For evaluation they recorded three SO-100 datasets (pick-and-place, stacking, sorting) and one SO-101 dataset (pick-and-place of a Lego brick into a transparent box), each with 50 demonstrations, and released them as lerobot/svla_so100_pickplace, svla_so100_stacking, svla_so100_sorting and svla_so101_pickplace.[16]

SettingACTπ0 (3.5B)SmolVLA (0.45B)
SO-100, average of three tasks (multi-task for π0 and SmolVLA, single-task for ACT)48.3%61.7%78.3%
SO-101 pick-place, in distribution (single-task)70%not reported90%
SO-101 pick-place, out of distribution (single-task)40%not reported50%

Expanded article table

The paper stresses that SmolVLA "is not pretrained on any datasets recorded for the SO101," so the SO-101 numbers are a transfer test to an embodiment absent from pretraining.[16]

Ecosystem

Kits. The README lists frame, electronics, parts and fully assembled kits from RobotEd (Switzerland), Robonine, PartaBot (US, also selling LeKiwi and Koch robots), ForgeMotion Labs (US), Seeed Studio (international, China and Japan), WowRobo, RoboSEasy (South Korea), NeoBot (China) and Autodiscovery (EU), plus an SO-100 follower-only kit from Phospho aimed at VR-headset teleoperation.[1] The launch partners named by Hugging Face were WowRobo, Seeed Studio and PartaBot.[3][7] Kit prices vary by vendor, level of assembly and country and are not tracked here.

LeKiwi. A low-cost mobile manipulator from SIGRobotics at the University of Illinois Urbana-Champaign that mounts an SO-ARM101 arm on a three-wheel holonomic base with omni wheels, driven by a Raspberry Pi 5 and powered either by a 12 V Li-ion pack or a 65 W laptop power bank. The repository (Apache 2.0) was created on 4 February 2025 and LeRobot has a dedicated LeKiwi page; the LeRobot paper puts its cost at about 230 EUR.[6][13]

XLeRobot. A dual-arm mobile home robot from the Vector-Wangel GitHub project that combines two SO-101 arms, a LeKiwi base, a 300 Wh Anker battery, two wrist RGB cameras and a head depth camera on a two-DoF neck, for a stated total of $660; it is listed in the SO-ARM100 README and its own repository (Apache 2.0, created 26 April 2025) had 5,526 stars on 16 September 2026.[1][21]

Bimanual and mobile setups. LeRobot added a bimanual SO-100 configuration in July 2025 and the SO-ARM100 README ships an overhead camera mount for "single or bi-manual setups."[1][17] Two of the 2026 papers below use bimanual SO-101 rigs.[19][22]

Simulation. The repository carries a URDF for the SO-100 and both URDF and MuJoCo MJCF files for the SO-101 (in "old" and "new" calibration variants), viewable with rerun.[15] NVIDIA's learning path uses those models in Isaac Sim to collect simulated demonstrations, apply domain randomisation and Cosmos augmentation, fine-tune GR00T, and deploy on a physical SO-101 for vial pick-and-place.[18]

Related Robot Studio hardware. The LeRobot paper also lists the HopeJR humanoid arm and hand from The Robot Studio (about 500 EUR) among supported platforms.[6]

Reception and use in research

Delangue called the SO-100 "the most popular robot arms ever?" at the SO-101 launch; the Hub statistics above are the closest thing to a measurable basis for that claim.[6][7] NVIDIA's learning path describes the SO-101 as "a 6-DOF (degrees of freedom) robot arm designed for research and education in manipulation tasks" and chose it for "first-class support in the LeRobot ecosystem" and available simulation models.[18]

Several 2026 papers use the arm as their sole platform:

  • ArmnetBench v0.1 (Selvaraj, Uttini and Kuosmanen, July 2026) runs a benchmark "on a fleet of low-cost SO-101 cells under light on-site supervision," comparing seven policies across twelve single-arm and bimanual tasks with 2,518 policy rollouts and 600 reference demonstrations, all released in LeRobot v3.0 format.[19]
  • Benchmarking Vision-Language-Action Models on SO-101 (Yu and Qiu, June 2026) fine-tunes π0.5, SmolVLA, Wall-X and ACT on teleoperated SO-101 demonstrations across four tasks and adds a failure taxonomy and recovery-aware metrics; it reports that pretrained VLA policies generally beat the ACT baseline.[20]
  • Bimanual Manipulation Within an 8 GB Budget (Singh et al., August 2026) runs a bimanual SO-101 system entirely on an NVIDIA Jetson Orin Nano Super, training ACT and diffusion policy on identical beanbag pick-and-place demonstrations; ACT reached 19 of 20 trials.[22]

The arm's limits follow from its parts. The servos are position-controlled hobby-class actuators, the structure is printed plastic, and the project publishes no payload, repeatability or safety rating, The 2026 benchmark paper frames the arm as a platform for studying policy robustness "under embodiment uncertainty" rather than as a precision manipulator.[20]

Comparison with other open arms

ArmOriginJointsActuatorsPayload (vendor)Reach (vendor)Indicative costSource
SO-101The Robot Studio and Hugging Face5 arm joints plus gripperFeetech STS3215 bus servos, 7.4 Vnot publishednot published$121.94 follower, $229.88 pair (project BOM, US, Sep 2026)[1][5]
Koch v1.1Jess Moss, revised from Alexander Koch's low-cost armleader and follower pairDynamixel servos (XL330-M077-T on the leader)not publishednot publishedabout 670 EUR per teleoperation setup (LeRobot paper)[6][23]
ALOHA (ViperX-based bimanual)Stanford (ALOHA), later ALOHA 2two arm pairsDynamixelsee ALOHA articlesee ALOHA articleabout 21,000 EUR (LeRobot paper)[6]
reBot Arm B601-RSSeeed Studio6 DOF plus gripperRobStride motors, 48 V2.5 kg754 mmnot stated in repository table[24]
OpenArm 2.0Enactic, Inc. (Tokyo)7 DOFDamiao motors, 9:1 and 10:1 backdrivable at shoulder and wrist, 40:1 (not QDD per Enactic) in the middle joints4.1 kg nominal, 6.0 kg peakhuman-scale (160-165 cm person)$6,500 for a complete bimanual system[25]

Expanded article table

The SO-101 sits at the bottom of this range on cost and on published capability. Its niche is volume: a printable design, sub-$250 parts for a full teleoperation pair, and a data pipeline that has produced thousands of public datasets, which is what the imitation learning and vision-language-action work above depends on.

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19TheRobotStudio, "Standard Open SO-100 & SO-101 Arms" (README, bill of materials, printing guide, optional hardware). GitHub. github.com/...SO-ARM100
  2. ^1 ^2 ^3 ^4GitHub repository metadata for TheRobotStudio/SO-ARM100 (created 10 May 2024, Apache-2.0, 7,475 stars and 685 forks on 16 September 2026). api.github.com/...SO-ARM100
  3. ^1 ^2 ^3TechCrunch, "Hugging Face releases a 3D-printed robotic arm starting at $100," TechCrunch, 28 April 2025. techcrunch.com/...nted-robotic-arm-starting-at-100
  4. ^1 ^2Hugging Face LeRobot documentation, "SO-100" (SO-100 assembly and configuration notes). huggingface.co/...so100
  5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Hugging Face LeRobot documentation, "SO-101" (assembly guide, leader gear-ratio table, motor setup and calibration). huggingface.co/...so101
  6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Remi Cadene et al., "LeRobot: An Open-Source Library for End-to-End Robot Learning," arXiv:2602.22818, 26 February 2026. arxiv.org/...2602.22818
  7. ^1 ^2 ^3 ^4Clement Delangue, post on X, 28 April 2025. x.com/...1916859453917241761
  8. ^1 ^2Remi Cadene, post on X, 28 April 2025. x.com/...1916751964807057515
  9. ^1 ^2TheRobotStudio/SO-ARM100, CITATION.cff. github.com/...CITATION.cff
  10. ^1 ^2 ^3TheRobotStudio/SO-ARM100, "Standard Open SO-100 Arm" (SO100.md, deprecated documentation). github.com/...SO100.md
  11. ^1 ^2 ^3 ^4 ^5TheRobotStudio/SO-ARM100, CHANGELOG.md. github.com/...CHANGELOG.md
  12. ^huggingface/lerobot, pull request #419 "Add FeetechMotorsBus, SO-100, Moss-v1," merged 25 October 2024. github.com/...419
  13. ^1 ^2SIGRobotics-UIUC, "LeKiwi - Low-Cost Mobile Manipulator" (README). GitHub. github.com/...LeKiwi
  14. ^TheRobotStudio/SO-ARM100, pull request #67 "Update" (commit message "Add SO101, cleanup repo"), merged 28 April 2025. github.com/...67
  15. ^1 ^2 ^3TheRobotStudio/SO-ARM100, "Simulation Models for SO100 and SO101" (Simulation/README.md). github.com/...README.md
  16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Mustafa Shukor, Dana Aubakirova, Francesco Capuano et al., "SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics," arXiv:2506.01844, 2 June 2025. arxiv.org/...2506.01844
  17. ^1 ^2huggingface/lerobot, pull request #1509 "Feat/add bimanual so100 robot," merged 16 July 2025. github.com/...1509
  18. ^1 ^2 ^3 ^4 ^5NVIDIA, "Train an SO-101 Robot From Sim-to-Real With NVIDIA Isaac," module "LeRobot: Background and Community." docs.nvidia.com/...04-lerobot
  19. ^1 ^2Praveen Selvaraj, Lorenzo Uttini and Ville Kuosmanen, "ArmnetBench v0.1: Parallel Real-World Evaluation of Manipulation Policies on a Low-Cost Arm Farm," arXiv:2607.24481, 27 July 2026. arxiv.org/...2607.24481
  20. ^1 ^2 ^3 ^4Yi Yu and Xinchuan Qiu, "Benchmarking Vision-Language-Action Models on SO-101: Failure and Recovery Analysis," arXiv:2606.08881, June 2026. arxiv.org/...2606.08881
  21. ^Vector-Wangel/XLeRobot, "XLeRobot: Practical Dual-Arm Mobile Home Robot for $660." GitHub. github.com/...XLeRobot
  22. ^1 ^2 ^3 ^4Ekansh Singh, Eva Samuel, Alessandra Reneau, Ryan Schmeelk and Yashvi Gandhi, "Bimanual Manipulation Within an 8 GB Budget: Zero-Copy Sensing and Quantized ACT on an Entry-Level Jetson," arXiv:2608.03938, 4 August 2026. arxiv.org/...2608.03938
  23. ^Jess Moss, "Low-Cost Robot Arm: Koch v1.1" (README). GitHub. github.com/...koch-v1-1
  24. ^Seeed-Projects/reBot-DevArm, "Hardware Specifications" table (README). GitHub. github.com/...reBot-DevArm
  25. ^Hugging Face LeRobot documentation, "OpenArm." huggingface.co/...openarm

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision · v2 · 3,153 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Adversarial verification 2026-09-16 (cluster V4): specs checked against the SO-ARM100 README, changelog and LeRobot docs; 2 material and 5 minor defects fixed

Cite this page: AI Wiki. "SO-101." aiwiki.ai, updated 16 Sept 2026, fact-checked 16 Sept 2026. CC BY 4.0. https://aiwiki.ai/wiki/so_101

Suggest edit