reBot Arm

reBot Arm is an open-hardware desktop robotic-arm series developed by Seeed Studio for robotics education, teleoperation, and embodied-AI experiments. The series has two main builds, reBot Arm B601-DM and reBot Arm B601-RS, which share a six-axis arm and one actuated gripper but use different motor families and electrical systems. Seeed publishes mechanical source files, bills of materials, control software, and integration guides for both builds.^[1][2]

The project is positioned as a development platform rather than a validated industrial robot. Its public materials document LeRobot, ROS, MoveIt, Pinocchio, NVIDIA Isaac Sim, and Isaac GR00T workflows. They do not provide an independent safety certification, service-life study, or controlled reliability benchmark for the arm.^[1][3]

Project history and licensing

Seeed created the public Seeed-Projects/reBot-DevArm repository in November 2025. The repository says the project initially used a noncommercial Creative Commons license, then changed its hardware license to CERN Open Hardware Licence Version 2, Weakly Reciprocal, on May 11, 2026. Software in the project is licensed under Apache License 2.0.^[1]

The Open Source Hardware Association certified reBot Arm B601 as open-source hardware on May 14, 2026, under identifier CN000024. The certification record covers the B601-DM and B601-RS, lists CERN-OHL-W-2.0 for hardware and documentation, and lists Apache for software.^[2]

The released hardware material includes STEP files for printed, machined, and purchased parts, along with bills of materials. Seeed cautions that the published B601-RS bill of materials does not necessarily represent the final shipping version. The open files therefore document a buildable design, but they should not be treated as an exact manufacturing record for every retail unit.^[4]

Hardware variants

Both B601 variants have six arm joints and an actuated gripper. The B601-DM uses Damiao motors and a 24 V supply. The B601-RS uses RobStride motors and a 48 V supply. Seeed's repository gives the following series-level specifications.^[1]

Current B601-RS documentation lists three RobStride RS06 motors and four RS00 motors across the six joints and gripper, without stating which joint carries which motor. The motors communicate over a 1 Mbps CAN bus. The same documentation lists a 2.5 kg rated payload, 5 kg maximum load, 754.7 mm reach including the gripper, and 587.5 mm reach without it.^[5] The 5 kg figure is a vendor-stated maximum rather than a rated continuous payload.

RobStride's own product pages classify both motor types as quasi-direct-drive ("QDD") integrated actuators, each combining a brushless motor, a planetary reducer with machined steel gears, a driver board and magnetic encoders in one housing. The RobStride 06 has a 9:1 reduction, 11 N m rated torque, 36 N m peak torque, a 48 V rated voltage and a listed weight of 621 g. The RobStride 00 has a 10:1 reduction, 5 N m rated torque, 14 N m peak torque, a 48 V rated voltage and a listed weight of 310 g.^[11][12] Seeed's Pinocchio guide reproduces the same reduction ratios and torque figures in its motor table.^[5] The B601-RS is therefore a quasi-direct-drive arm with 9:1 and 10:1 reductions, not a direct-drive arm in the strict sense of motors coupled to the joints without any reduction. Third-party summaries that call it "direct-drive" are using the term loosely.

For the B601-DM, Seeed published a performance-testing reference on April 1, 2026, based on a build with Damiao V4 motors. In reciprocating motion over 5 to 70 percent of reach with a 1.5 kg load, the arm ran for more than two hours before the second joint motor reached 90 degrees C and the test was stopped manually; with 2.5 kg the motor's overheat protection tripped after 40 minutes, and with 1.5 kg over the full reach it tripped after 45 minutes. Static holds of 1.5 kg lasted 18 minutes at 70 percent extension and 3 minutes at full extension. Seeed concludes that the structure is strong enough and that the limit is motor heating, and it recommends loads of no more than 1.5 kg, a working radius under 70 percent of reach (about 450 mm), and active cooling for sustained high-load use.^[14] The document states that the figures are for reference only and differ for earlier motor versions.

Seeed's first-party pages disagree on two B601-RS measurements. The September 2026 course states less than 0.1 mm (and plus or minus 0.1 mm in its prose) and the B601-RS Pinocchio guide 0.1 mm, while the main repository table states less than 0.2 mm. The course and the Seeed store page list a 6.5 kg arm weight, whereas the repository and Pinocchio guide list approximately 6.7 kg.^{[1][3][5][13]} Seeed's retail listings add a third phrasing: the B601-RS product page advertises "±0.1 mm repeatability" and the B601-DM bundle page "±0.2 mm".^[13][25] Because the sources do not explain whether these differences reflect measurement methods or a hardware revision, the values should not be combined into a single definitive specification.

Control and software stack

The reBot repositories separate low-level motor access from higher-level robotics tools. The Python control layer handles motor communication and arm configuration. Seeed supplies guides for kinematics and gravity compensation with Pinocchio, motion planning through ROS and MoveIt, camera-based grasping examples, LeRobot data collection, and USD-based simulation.^[1][6]

NVIDIA's Isaac Teleop documentation includes a rebot_devarm_leader plugin for the seven reBot joints. The plugin selects the Damiao backend when given a serial-device path and the RobStride backend when given a SocketCAN interface such as can0. NVIDIA documents the Damiao build as seven MIT-protocol motors behind a USB-to-CAN adapter. The RobStride path uses classic CAN at 1 Mbps and is Linux-only. Both paths can disable motor torque so an operator can back-drive the arm while joint positions are read for teleoperation.^[7]

For B601-DM, Seeed also publishes an Isaac Sim bridge for synchronizing a physical or manually driven arm with a simulated model. Its documented modes include joint tests without physical hardware, inverse kinematics, minimum-jerk joint-space trajectories, gravity-compensated manual input, and passive real-to-sim joint mirroring. In the documented setup, sender and receiver processes exchange joint data over UDP.^[6]

The B601-DM is also supported directly in the upstream LeRobot library. A pull request titled "natively integrate Seeed Studio reBot B601-DM arm" was merged on May 18, 2026, adding a rebot_b601_follower robot class, a bimanual bi_rebot_b601_follower class, and a documentation page; an MIT-mode control option followed on June 30, 2026. The LeRobot page describes the follower as a 6-DOF arm plus gripper driven by Damiao CAN motors through the motorbridge package, paired with a Star Arm 102 leader built on FashionStar UART servos, and it accepts either a Damiao serial bridge or a SocketCAN adapter.^[15][16] The upstream integration is specific to the Damiao build. Seeed's LeRobot guide for the B601-RS instead has users clone Seeed's own fork of LeRobot.^[17]

NVIDIA's official Isaac-GR00T repository contains a separate B601-DM example. It converts a LeRobot dataset, fine-tunes NVIDIA Isaac GR00T N1.7 as a new embodiment, performs open-loop evaluation, and provides a closed-loop client example for the physical arm.^[8] This published example is specific to the Damiao build. The B601-RS course described below supplies its own RS-specific configuration rather than establishing that every reBot configuration is interchangeable.

Sim-to-real VLA course

On September 1, 2026, NVIDIA Robotics announced a course developed by Seeed Studio in collaboration with NVIDIA's robotics team. The course, titled Learning Physical AI: A Sim-to-Real VLA Pipeline with Seeed reBot Arm and NVIDIA Isaac, uses the B601-RS as its physical follower arm. Its public interface lists 19 modules in five chapters and more than 20 hours of intermediate-level material.^[3][9]

The course organizes a sim-to-real transfer workflow around a stationery pick-and-place task. It covers physical and simulated demonstration collection, optional synthetic scene augmentation, post-training of a vision-language-action model, real-arm evaluation, and edge deployment. These are tutorial stages, not evidence that one trained policy will transfer to every task or environment.^[3]

Demonstration collection

For physical data collection, the course uses a Star Arm 102 leader arm to command a B601-RS follower through LeRobot. Its example records two 640 by 480 camera streams at 30 frames per second, joint and gripper data, and a language task label. The sample command requests 100 episodes of a 25-second stationery-organization task. Those values define the tutorial example and are not minimum dataset requirements established by a controlled study.^[3]

The simulated-data module is explicitly optional. It uses a reBot leader to control a B601 model in Isaac Sim and stores actions, simulated joint states, wrist and side-camera images, task text, randomization parameters, and task-success status in LeRobot Dataset v3 format. The module also supports replaying recorded actions in simulation. Replay checks the motion trace; it is not an evaluation of physical transfer.^[3]

Augmentation and GR00T post-training

An optional augmentation module uses NVIDIA Cosmos 3 Transfer to change scene appearance while retaining the source manipulation video. The course combines a source video with Canny edges, SAM 2 segmentation, and a text description of the target background. It tells learners to inspect generated videos for robot deformation, object inconsistency, temporal flicker, and viewpoint errors before adding them to training data.^[3]

The training module converts its LeRobot v3 data to the v2 layout expected by the illustrated GR00T workflow, maps six arm joints and the gripper into a seven-dimensional state and action representation, and fine-tunes GR00T N1.7 as a new embodiment. GR00T's published research describes the model family as a dual-system VLA architecture: a vision-language component interprets images and instructions, and a diffusion-transformer component generates actions.^[10] That paper evaluates GR00T N1 across other embodiments. It does not report results for reBot, so its benchmark claims cannot be transferred to the B601 series.

Physical evaluation and edge deployment

The course's physical-evaluation module separates model inference from robot control through a policy server and client. Its demonstration reports autonomous manipulation by the trained arm, but the page does not publish episode counts, success rates, comparison baselines, failure categories, or statistical uncertainty.^[3]

The deployment chapter describes exporting the trained policy to TensorRT for NVIDIA Jetson systems. Its example builds seven target-specific engines and warns that engine files cannot be copied between Jetson AGX Orin and Jetson Thor. Engine compatibility depends on the GPU architecture, TensorRT version, checkpoint, graph shapes, precision, and builder configuration.^[3]

Comparison with other open-source arms

The reBot B601 sits between the cheapest 3D-printed hobby arms and the larger human-scale research arms. The table below compares it with three arms that are widely used for imitation-learning and teleoperation work: the SO-101, OpenArm 2.0, and the ViperX-300 arms used in the ALOHA bimanual system. Every cell comes from the respective project's own documentation; where a project publishes no figure, the cell says so.

Several details in the table deserve comment. The SO-101 is often described as a six-DOF arm, but its six servos drive five arm joints (shoulder pan, shoulder lift, elbow flex, wrist flex, wrist roll) and the gripper, so its arm has one fewer joint than the B601.^[28] OpenArm's payload definitions are unusually explicit (a one-minute hold in the worst posture for the nominal figure, a three-second lift plus one-second hold for the peak figure) and include the end effector, whereas Seeed's specification table and Pinocchio guide give the B601-RS figure of 2.5 kg without describing a test protocol.^[1][5][19] OpenArm also says plainly that one of its motor families is not quasi-direct-drive, so neither OpenArm nor the B601-RS is a uniformly "direct-drive" arm.^[20] The ViperX figure is a working payload with an accuracy of 5 to 8 mm and 1 mm repeatability according to Trossen, which is a different accuracy class from the sub-millimetre repeatability Seeed claims.^[22] Finally, the reBot B601 carries an OSHWA certification (CN000024); the SO-101 and OpenArm projects publish their hardware files under open licenses on GitHub, and the ViperX is a commercial product, with the ALOHA paper open-sourcing the surrounding frame, 3D-printed fingers and software rather than the arm itself.^{[2][18][21][23]}

Availability and recent developments

Seeed sells both builds through its own store. On September 16, 2026, the assembled B601-RS with gripper (SKU 100019336, sold without a power supply) was listed at $1,799.99 and shown as in stock.^[13] The B601-DM is sold as separate kits: a body motor kit at $994.80, a body structure kit at $202.80, and a gripper kit at $238.80 on the same date, with a leader-arm kit and power and emergency-stop accessories offered alongside.^[25] The repository also links an AliExpress listing and a Star Arm 102-LD leader arm from FashionStar, a link added on September 3, 2026.^[1] Prices on Seeed's store change without notice, so these figures are a snapshot.

Repository activity after the September 1 course launch has been mostly documentation. On September 7, 2026, the README gained a link (committed the day before) to a browser-based MuJoCo digital twin of the B601-RS, hosted on a contributor's GitHub Pages site, which lets users drive the arm's joints and tool centre point, switch between a fixed-base and a mobile (AGV) configuration, view simulated overhead and wrist cameras, and run scripted put-away and stacking demonstrations.^[1][27] On September 7, the README added an "Embodied Agent Architecture" tutorial for the B601-RS: a demo that takes natural-language instructions such as "pick up the red block", uses a vision-language model (Qwen3-VL-2B by default, with cloud model options) to select objects, solves inverse kinematics with Pinocchio, and passes every 50 Hz waypoint through a fail-closed safety check before execution.^[1][26] Community-contributed design notes for a RealSense D405 wrist mount were restored to the README in a change merged on September 11.^[1] As of September 16, 2026, the repository had 4,252 stars and 427 forks and had published no tagged releases.^[1]

Limitations

Most detailed evidence about reBot comes from Seeed Studio, its repositories, and documentation produced with NVIDIA. The NVIDIA Isaac Teleop integration and Isaac-GR00T example independently confirm that the arm has concrete integration paths in NVIDIA's software, but they do not establish mechanical durability, payload performance over time, calibration stability, or general policy reliability.^[7][8] The only published endurance data is Seeed's own April 2026 test of the Damiao build, which was limited by motor heating rather than structure and which Seeed itself labels as reference-only.^[14]

The September course is a reproducible tutorial rather than a peer-reviewed benchmark. It demonstrates one stationery task and gives no controlled comparison between real-only, simulation-assisted, and Cosmos-augmented training. It also does not quantify the sim-to-real gap or report how often the physical task succeeds. Claims that the workflow improves robustness or reduces data requirements therefore remain unverified for reBot.^[3]

Users must also reconcile version-specific dependencies. The course pins particular Isaac Sim, LeRobot, GR00T, camera, and Jetson environments, while the standalone reBot documentation continues to evolve. Hardware limits, motor configuration, joint calibration, camera mapping, CAN setup, inference latency, and emergency-stop procedures remain part of deployment. A successful simulation or vendor demonstration is not a substitute for task-specific safety validation on the physical arm.^[3][5][6]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Seeed Studio. "Seeed-Projects/reBot-DevArm." GitHub repository and README. Accessed September 1, 2026. github.com/...reBot-DevArm
2. ^1 ^2 ^3Open Source Hardware Association. "reBot Arm B601," certification CN000024. Certified May 14, 2026. certification.oshwa.org/cn000024
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Seeed Studio and NVIDIA Robotics. "Learning Physical AI: A Sim-to-Real VLA Pipeline with Seeed reBot Arm and NVIDIA Isaac." Course modules. Accessed September 1, 2026. seeedstudio.com/...th-seeed-rebot-and-nvidia-isaac
4. ^Seeed Studio. "reBot DevArm Open Source Hardware Specification: reBot Arm B601 RS." GitHub. Initial hardware upload dated July 9, 2026. github.com/...README.md
5. ^1 ^2 ^3 ^4 ^5 ^6Seeed Studio. "Getting Started with Pinocchio and MeshCat for reBot Arm B601-RS." Accessed September 1, 2026. wiki.seeedstudio.com/..._b601_rs_pinocchio_meshcat
6. ^1 ^2 ^3Seeed Studio. "Simulating reBot-B601-DM with Isaac Sim." Accessed September 1, 2026. wiki.seeedstudio.com/rebot_arm_b601_dm_isaacsim
7. ^1 ^2NVIDIA. "Generic Joint-Space Device: The reBot DevArm leader plugin." Isaac Teleop 1.4 documentation. Updated August 21, 2026. nvidia.github.io/...joint_space
8. ^1 ^2 ^3NVIDIA. "Finetuning GR00T N1.7 Model for Seeed reBot Arm B601 DM." Isaac-GR00T repository. Accessed September 1, 2026. github.com/...README.md
9. ^NVIDIA Robotics. X post announcing the Seeed reBot Arm and NVIDIA Isaac course. September 1, 2026. x.com/...2094850247184826558
10. ^NVIDIA et al. "GR00T N1: An Open Foundation Model for Generalist Humanoid Robots." arXiv:2503.14734, revised March 27, 2025. arxiv.org/...2503.14734
11. ^1 ^2RobStride. "RobStride 06" product page (QDD 36 N m integrated actuator; 9:1 reduction ratio; 11 N m rated and 36 N m peak torque; 621 g). Accessed September 16, 2026. robstride.com/...robStride06
12. ^1 ^2RobStride. "RobStride 00" product page (QDD 14 N m integrated actuator; 10:1 reduction ratio; 5 N m rated and 14 N m peak torque; 310 g). Accessed September 16, 2026. robstride.com/...robStride00
13. ^1 ^2 ^3Seeed Studio. "reBot Arm B601-RS (without power supply)," SKU 100019336, product page with price and stock status. Accessed September 16, 2026. seeedstudio.com/...sembled-Kit-with-Gripper-p-6865
14. ^1 ^2Seeed Studio. "reBot-DevArm Real Machine Performance Testing Reference," v1.0, April 1, 2026. github.com/...Performance_Testing.md
15. ^1 ^2Hugging Face. "reBot B601-DM." LeRobot documentation. Accessed September 16, 2026. huggingface.co/...rebot_b601
16. ^huggingface/lerobot. Pull request #3624, "feat(robots): natively integrate Seeed Studio reBot B601-DM arm," merged May 18, 2026. github.com/...3624
17. ^Seeed Studio. "Getting Started with reBot Arm B601-RS in LeRobot." Accessed September 16, 2026. wiki.seeedstudio.com/rebot_arm_b601_rs_lerobot
18. ^1 ^2TheRobotStudio. "Standard Open SO-100 & SO-101 Arms," README with bill of materials (Apache-2.0 repository). Accessed September 16, 2026. github.com/...SO-ARM100
19. ^1 ^2Enactic. "OpenArm 2.0: General" (payload definitions, key-features graphic). Accessed September 16, 2026. docs.openarm.dev/...general
20. ^1 ^2Enactic. "OpenArm 2.0: Motor" (motor selection and specification table). Accessed September 16, 2026. docs.openarm.dev/...motor
21. ^1 ^2Enactic. "enactic/openarm" README (repository and license table, $6,500 bimanual system). Accessed September 16, 2026. github.com/...openarm
22. ^1 ^2Trossen Robotics. "ViperX-300 6DOF," Interbotix X-Series Arms Documentation (6 DoF, 750 mm reach, 1,500 mm span, 1 mm repeatability, 5-8 mm accuracy, 750 g working payload, 9 servos). Accessed September 16, 2026. docs.trossenrobotics.com/...vx300s
23. ^1 ^2Tony Z. Zhao, Vikash Kumar, Sergey Levine and Chelsea Finn. "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware." arXiv:2304.13705, 2023. arxiv.org/...2304.13705
24. ^tonyzhaozh/aloha. "ALOHA: A Low-cost Open-source Hardware System for Bimanual Teleoperation," repository README (MIT license; ROS 1 Noetic and Interbotix setup). Accessed September 16, 2026. github.com/...aloha
25. ^1 ^2Seeed Studio. "reBot Arm B601-DM Unassembled Bundle," product page listing the body motor, body structure and gripper kits. Accessed September 16, 2026. seeedstudio.com/reBot-Arm-B601-DM-Bundle
26. ^Seeed Studio. "Embodied Agent Architecture Design for reBot B601 RS Robotic Arm." Accessed September 16, 2026. wiki.seeedstudio.com/wrc_demo_tutorial
27. ^Yang-Ci. "B601-RS-Sim," browser-based MuJoCo digital twin linked from the reBot-DevArm README. Accessed September 16, 2026. yang-ci.github.io/Rebot_Arm_AGV
28. ^1 ^2huggingface/lerobot. `so_follower.py` (motor table: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper). Accessed September 16, 2026. github.com/...so_follower.py