Overview

Three arms dominate manipulation research in 2025: the OpenArm 101, the Franka Emika Panda, and the Kinova Gen3. They cover a price range from $8K to $40K, have fundamentally different software philosophies, and are optimized for different use cases. Choosing the wrong one costs more than money — it costs months of integration work.

Quick Specs Comparison

SpecOpenArm 101Franka PandaKinova Gen3
DOF777
Payload1.5 kg3 kg4 kg
Reach900 mm855 mm902 mm
Repeatability±0.5 mm±0.1 mm±0.1 mm
Price (USD)~$8,000$25K used / $40K new~$30,000
Primary SDKPython / SocketCANFCI (C++ / Python)Kortex API (Python/C++)
Force-torque sensingOptionalBuilt-in (joint torque)Optional wrist
ROS2 supportNativeros2_franka (community)Official
IP ratingIP40IP40IP54
Open hardwareYesNoNo

OpenArm 101: The Community-Driven Option

OpenArm 101 is a 7-DOF arm designed specifically for manipulation research and imitation learning. At roughly $8,000, it is the most accessible serious research arm on the market. The design files are fully open-source, meaning labs can modify the end-effector, add custom sensors, or manufacture replacement parts without vendor lock-in.

Its primary strengths: SVRC platform integration (OpenArm is the reference hardware for the SVRC data collection stack — teleoperation, data logging, and replay all work out of the box), SocketCAN communication (direct, low-level joint control at 1kHz update rates), and community (the SVRC community forum has active OpenArm threads, shared configurations, and a growing library of task setups).

Its primary weakness is repeatability (±0.5mm vs. ±0.1mm for Franka and Kinova). For most imitation learning tasks this difference is irrelevant — policies operate in closed-loop and can correct for 0.5mm errors. For precision assembly with sub-millimeter tolerances, the repeatability gap matters.

Franka Emika Panda: The Research Standard

The Franka Panda is the academic standard for a reason. Hundreds of manipulation papers have used it, meaning there is an enormous base of code, configurations, and reference results. Its built-in joint torque sensing enables force-controlled manipulation out of the box — gravity compensation, compliance control, and contact detection all work without additional hardware.

The Franka Cartesian Interface (FCI) gives direct 1kHz control access, which is important for researchers implementing custom control algorithms. MoveIt2 integration is mature. The robot is reliable enough that many labs run it continuously for months without mechanical issues.

The cost ($25K used, $40K new) and the Franka-specific software ecosystem are the main downsides. If your lab already has a ROS2 Franka stack and existing data, staying with Franka is the right call — the switching cost to another platform is higher than the price difference.

Kinova Gen3: Mobile Platform Compatibility

The Kinova Gen3 is differentiated by two things: IP54 weather/dust rating and its design for mounting on mobile bases. The IP54 rating makes it suitable for environments where robot wash-down or outdoor operation is required. Its weight distribution is optimized for mobile platforms — Kinova publishes verified configurations for Hello Robot Stretch and several AMR platforms.

The integrated depth camera (on Gen3 Vision configuration) is convenient for labs that want an out-of-box vision stack without additional calibration work. The Kortex API is well-documented and has Python bindings that are more accessible than Franka's FCI.

For pure manipulation research on a fixed bench, Kinova is harder to recommend over Franka — the repeatability is comparable but the research ecosystem is smaller. The value proposition is strongest for mobile manipulation research.

Full Specifications Comparison

SpecificationOpenArm 101Franka PandaKinova Gen3
DOF777
Payload1.5 kg3 kg4 kg
Reach900 mm855 mm902 mm
Repeatability+/-0.5 mm+/-0.1 mm+/-0.1 mm
Max joint speed180 deg/s150 deg/s (J1-J4), 180 (J5-J7)150 deg/s
Max end-effector speed1.5 m/s2.0 m/s0.5 m/s (spec), 1.0 m/s (achievable)
Weight8 kg18 kg8.2 kg
Force sensingOptional wrist F/TBuilt-in joint torque sensorsOptional wrist F/T
CommunicationSocketCAN / USBFCI (Ethernet, 1kHz)Ethernet / USB / WiFi
Control rate1 kHz1 kHz1 kHz (low-level API)
OS supportLinux (Ubuntu 22.04+)Linux (Ubuntu 20.04/22.04)Linux, Windows
ROS2 supportNative (maintained)ros2_franka (community)Official ros2_kortex
LeRobot supportCommunity adapter (SVRC)Community adapterNot yet
IP ratingIP40IP40IP54
Open hardwareYes (full CAD + BOM)NoNo
Price (USD)~$4,500$25K used / $40K new~$30,000

Software Ecosystem Depth

OpenArm 101: Python SDK is the primary interface -- clean, well-documented, and accessible to ML researchers who may not have C++ experience. SocketCAN provides raw CAN bus access for teams that need low-level control. The SVRC data platform provides end-to-end data collection, recording, and training integration. LeRobot support via community adapter. URDF model available for MoveIt2 and simulation. Growing library of task configurations and shared datasets on the SVRC community forum.

Franka Panda: The deepest software ecosystem of any research arm. Hundreds of published papers with accompanying code, most using the FCI (C++) or franka_ros (ROS1/ROS2) interface. MoveIt2 integration is the most mature of any arm. libfranka provides millisecond-precise Cartesian and joint impedance control. The Drake simulator includes a validated Franka model. Downside: the FCI API has a steep learning curve for teams without C++ experience, and the real-time Linux requirement adds system administration overhead.

Kinova Gen3: The Kortex API is the most user-friendly of the three -- clean Python bindings, well-documented REST API for high-level control, and official ROS2 support through ros2_kortex. The integrated web interface allows basic operation without writing any code. For teams coming from a non-robotics background (e.g., ML teams adding a robot to their lab), Kinova has the lowest software learning curve. The ecosystem is smaller than Franka's but growing, particularly for mobile manipulation research.

Learning Curve and Community Support

FactorOpenArm 101Franka PandaKinova Gen3
Time to first demo2-4 hours1-3 days4-8 hours
Time to production data collection1-2 weeks2-4 weeks1-3 weeks
Community sizeGrowing (SVRC forum)Large (hundreds of labs)Medium
Published papers using it~20500+~80
Repair/maintenanceUser-serviceable (open hw)Vendor service requiredVendor service recommended

Use Case Matrix: Which Arm for Which Task

Use CaseBest ChoiceWhy
IL data collection (budget)OpenArm 1015x cheaper; SVRC platform integration
Precision assembly researchFranka Panda+/-0.1mm repeatability; built-in torque sensing
Mobile manipulationKinova Gen3IP54; lightweight; mobile platform configs
Multi-arm research fleetOpenArm 1014 OpenArms for the price of 1 Franka
Force-controlled insertionFranka PandaNative impedance control; joint torque sensing
Classroom / teachingOpenArm 101Low cost; open hardware for learning
Outdoor / dusty environmentsKinova Gen3Only IP54-rated option

Which Should You Choose?

  • Starting an imitation learning project with limited budget: OpenArm 101. The SVRC integration, open hardware, and community support make data collection and policy training faster to set up than any alternative at this price.
  • Lab with existing ROS2 Franka stack or needing sub-mm precision: Stay with or buy Franka. The ecosystem depth and force-control capabilities are worth the price premium for precision research.
  • Mobile manipulation research: Kinova Gen3. The IP54 rating and mobile platform compatibility are unique in this class.
  • Budget is the primary constraint: OpenArm 101 is the only sub-$10K option with serious research credentials.

You can browse OpenArm 101 configurations and accessories in the SVRC store. For questions about which arm fits your specific research agenda, the community forum has active discussion on hardware selection.

Related Reading