OpenArm vs SO-101: Which Robot Arm for Robot Learning? (2025)

A detailed side-by-side comparison of the two most popular low-cost open-source robot arms for imitation learning and data collection research.

Intent

Buying Decision

Use this when SO-101 and OpenArm are both on your shortlist and you need to commit to one.

Audience

Researchers & ML Engineers

Designed for teams who want to collect imitation learning data, not industrial automation.

Output

Clear Recommendation

Concrete guidance on which arm fits your task type, budget, and timeline.

Both the OpenArm 101 and the SO-101 (formerly SO-ARM100) are low-cost, open-source robot arms designed for robot learning research. Both support LeRobot. Both can be assembled in a weekend and driven by a laptop. But they make different engineering trade-offs that matter significantly depending on your task type, operator experience, and data collection goals. This guide breaks them down completely.

Specs Side-by-Side

Specification	OpenArm 101	SO-101
Degrees of freedom	6 DoF arm + gripper	6 DoF arm + gripper
Actuator type	MIT-style quasi-direct drive motors (Unitree / custom)	Dynamixel servo motors (STS3215 or similar)
Payload capacity	~1 kg at full extension	~0.5 kg at full extension
Reach (max extension)	~650 mm	~500 mm
Repeatability	±1–2 mm	±2–4 mm
Control interface	CAN bus (SocketCAN), USB-CAN adapter	USB serial (Dynamixel U2D2 or similar)
Control frequency	Up to 500 Hz (MIT control), 50 Hz typical for policy inference	Up to 50 Hz (Dynamixel protocol)
Torque sensing	Yes — joint-level torque feedback from quasi-direct drive	No — position control only
Price (single arm)	~$800–1,200 (kit from SVRC store)	~$350–500 (self-sourced) / ~$500–700 (assembled kit)
Bimanual configuration	Yes — DK1 bimanual kit (2× OpenArm + base plate)	Yes — SO-101 bimanual kit (2× arms)
Teleoperation method	Leader-follower (OpenArm leader arm), VR, keyboard	Leader-follower (SO-101 leader arm), keyboard
LeRobot support	Yes — official HuggingFace LeRobot integration	Yes — first-class support (designed for LeRobot)
ROS 2 support	Yes — ros2_control package, MoveIt2 compatible	Limited — community-maintained URDF, no official ros2_control
URDF / simulation	Yes — Isaac Sim, Mujoco, PyBullet URDFs available	Yes — Mujoco and PyBullet URDFs available
Assembly time	2–4 hours (pre-assembled option available)	3–6 hours (manual Dynamixel servo horn assembly required)
Commercial availability	SVRC store (US stock, ships within 2 business days)	HuggingFace / various resellers; longer lead times outside China
License	Apache 2.0 (hardware design + software)	Apache 2.0 (hardware design + software)

The Key Mechanical Difference: Actuators

The most important difference between these two arms is not the kinematics — it is the actuator technology. OpenArm uses quasi-direct drive (QDD) motors: high-torque, low-gear-ratio actuators with integrated encoders and torque sensing. SO-101 uses Dynamixel servos: high-gear-ratio position servos widely used in hobby and research robotics since the early 2000s.

What Quasi-Direct Drive Gives You (OpenArm)

Torque feedback: Every joint reports torque in real time. This enables compliant control, contact detection, and force-sensitive manipulation strategies.
Backdrivability: You can physically move the arm by hand when in compliance mode. This makes kinesthetic teaching and physical guidance possible — you hold the arm and move it, and the trajectory is recorded.
Higher bandwidth: CAN bus at 500 Hz enables much faster control loops than Dynamixel at 50 Hz. This matters for dynamic tasks and contact-rich manipulation.
Lower gear ratio = less position accuracy but more transparency: You feel real mechanical impedance, not a hidden gear train absorbing all interaction forces.

What Dynamixel Servos Give You (SO-101)

Simpler software stack: Dynamixel Protocol 2.0 is extremely well-documented. There are Python libraries, ROS packages, and community tutorials going back years.
Lower cost and sourcing ease: Dynamixel STS3215 servos are available globally, with multiple suppliers. If one burns out, you replace a $15–30 servo, not a $100–200 QDD motor.
Sufficient for LeRobot ACT workflows: For leader-follower teleoperation with ACT or Diffusion Policy training, 50 Hz position control is sufficient. Most successful SO-101 papers and videos use exactly this setup.
Lightweight and compact: SO-101 at ~800g is easier to mount, reposition, and transport. OpenArm's QDD motors add weight and bulk.

Setup Complexity Comparison

Both arms are designed to be assembled by researchers without specialized mechanical engineering skills. But the setup experience differs:

Step	OpenArm 101	SO-101
Physical assembly	Bolt-together aluminum links; QDD motors pre-wired. ~2 hrs.	3D-printed / aluminum links; Dynamixel horn installation requires care. ~3–5 hrs.
Driver / SDK install	SocketCAN setup, `openarm-sdk` pip install. Documented. ~30 min.	U2D2 USB adapter, `lerobot` pip install. Simpler. ~15 min.
Calibration	Joint homing + MIT control gain tuning required. 1–2 hours first time.	Dynamixel calibration script. ~20 minutes.
First teleoperation	Same day (SVRC pre-assembled option) or day 2 (self-assembled).	Day 1–2 depending on assembly speed.
Common failure modes	CAN bus wiring issues; motor homing failures. Well-documented in SVRC guides.	Dynamixel servo overheating on long sessions; USB COM port conflicts on Windows.
Support quality	SVRC direct support, same-day response. On-site help available in Mountain View.	Community support (HuggingFace Discord, GitHub issues). Response time varies.

LeRobot Compatibility

Both arms have first-class LeRobot support — but the integration depth differs.

SO-101 + LeRobot is the original design target. Remi Cadene and the HuggingFace robotics team designed SO-ARM100/SO-101 in parallel with LeRobot. The lerobot repository ships with native SO-101 motor configuration files, calibration scripts, and example teleoperation notebooks. If you clone LeRobot and have a SO-101, you can run teleoperation data collection within minutes of setup. The dataset format (HDF5 + metadata JSON) is natively compatible with ACT and Diffusion Policy training pipelines.

OpenArm + LeRobot is fully supported but requires a thin adapter layer. The OpenArm SDK exposes a lerobot-compatible interface — you instantiate the arm as a LeRobot environment and the teleoperation, recording, and dataset creation pipelines work identically to SO-101. SVRC maintains the integration layer and tests it against LeRobot main branch with each OpenArm firmware release. The primary difference is latency: OpenArm's CAN interface enables true 50 Hz synchronous control with lower jitter than SO-101's USB serial, which is important for the temporal consistency of recorded demonstrations.

Data Collection Performance

For imitation learning, the quality of your demonstration dataset is the primary driver of policy performance. Here is how the two arms compare on the dimensions that matter most:

Trajectory Smoothness

OpenArm's QDD motors produce smoother trajectories in leader-follower teleoperation because backdrivability allows the human operator to feel resistance and adjust naturally. When you push on the leader arm, you feel the payload — this feedback loop produces demonstrations with more consistent velocities and fewer jerky transitions. SO-101's geared servos do not backdriving naturally; the leader arm is effectively a duplicate arm running in position control, which can produce more mechanical-feeling demonstrations.

In practice: OpenArm demonstrations tend to have lower action variance within consistent task executions, which helps ACT's chunk prediction. SO-101 demonstrations are still high-quality for tasks where smooth trajectories matter less (pick-and-place, cup stacking).

Episode Consistency

Both arms are capable of collecting 50–200 episode datasets that train well. The SO-101 bimanual community has published impressive results (80%+ success on ALOHA-inspired tasks) with the arms. OpenArm's higher repeatability (±1–2mm vs ±2–4mm) matters most for tasks requiring precise placement — peg insertion, slot alignment, connector mating.

Episode Length

Both arms support episodes up to 30+ seconds without issues. OpenArm's Dynamixel servo overheating is not a concern (QDD motors run cooler on sustained sessions). SO-101 with STS3215 servos can overheat during long continuous-duty sessions — implement 5-minute rest periods if running 100+ episode sessions in a single sitting.

Demonstrated Task Range

SO-101: cup stacking, block sorting, shirt folding (lighter objects), object transfer, pen picking. Published success rates range 70–95% on standard tasks with 50–100 demos and ACT.

OpenArm: all SO-101 tasks plus heavier object manipulation (0.5–1 kg), contact-rich insertion tasks, and tasks requiring force modulation. The DK1 bimanual configuration additionally supports cloth manipulation, bimanual assembly, and whole-body-adjacent tasks.

Community and Support

SO-101 has the larger community as of early 2025. It was among the first arms to get official LeRobot support, and the HuggingFace Discord has hundreds of SO-101 users. There are multiple YouTube build guides, community-contributed fixes for common hardware issues, and a growing dataset library on HuggingFace Hub from SO-101 users. If you get stuck, someone has probably already hit the same issue and documented the fix.

OpenArm has a smaller but growing community supported directly by SVRC. The advantage is support quality: SVRC provides direct email and Discord support, with same-day response for hardware issues. For US-based labs, SVRC offers on-site support visits in the Bay Area and Boston. OpenArm ships with detailed SVRC documentation — calibration, ROS 2 control, MIT gains, SocketCAN setup — that is more comprehensive than what the SO-101 community has assembled. OpenArm also has an active issues tracker and monthly firmware updates.

Who Should Buy Each

Buy OpenArm 101 if:

You are running contact-rich manipulation tasks, need torque sensing, want to explore compliant control, plan to scale to bimanual with DK1, are in the US and want same-day support, or have tasks requiring better than 2mm repeatability. OpenArm is the stronger platform for research that will push beyond simple pick-and-place into force-sensitive or dexterous manipulation.

Buy SO-101 if:

You are starting from scratch with LeRobot and want the fastest on-ramp, have a tight budget (under $500 per arm), are running standard tabletop manipulation tasks (cup stacking, object sorting, block transfer), or want the largest community base for troubleshooting. The SO-101 is the better choice for teams new to robot learning who want to get demos recorded and a policy trained in the first week.

For Bimanual: Both Work, but DK1 Has an Edge

Both arms have bimanual configurations. The SO-101 bimanual kit (two arms + mounting) is ~$800–1,200 total and is the cheapest path to bimanual robot learning in existence. The DK1 (two OpenArm 101s on a structured base) runs ~$2,000–2,500 but adds torque sensing on both arms, higher payload, and the structural rigidity needed for symmetric bimanual tasks. If your research goal is ALOHA-style bimanual manipulation at scale, DK1 is the more capable platform. If you want to explore bimanual on a student budget, SO-101 bimanual is remarkable for the price.

Frequently Asked Questions

Can I run the same LeRobot ACT training script on both?

Yes. Both arms output LeRobot-compatible HDF5 datasets. The training command is identical. The only difference is the robot configuration YAML you pass to the data collection script.

Is SO-101 data transferable to OpenArm or vice versa?

Not directly — the kinematics differ, so action sequences from one arm do not replay on the other. But you can train policies on data from either arm and evaluate on the same arm. Cross-arm policy transfer is an open research problem.

Which arm is better for a first-year PhD student?

SO-101 for pure imitation learning research — simpler setup, larger community, faster first results. OpenArm if the research involves contact, force, or compliance from day one.

Does SVRC ship SO-101?

SVRC stocks and ships OpenArm 101 and DK1. For SO-101, we recommend the HuggingFace official store or authorized resellers. We support both arms in our teleoperation and data collection services.

Can I upgrade from SO-101 to OpenArm later?

Yes. Your LeRobot training scripts and data pipelines are compatible with both. The hardware upgrade is a full swap, but your software investment transfers completely.

Guide