Franka Research 3 (FR3)

Home > Hardware > Arms > Franka Research 3

Overview

Franka Research 3 (FR3) is a next-generation 7-axis collaborative robot developed by Franka Robotics (formerly Franka Emika) in Germany, an evolved version of the original Franka Panda. With torque sensors embedded in all joints and support for 1kHz real-time control, it has established itself as the de facto Reference Platform for AI and robotics research. As of February 2025, more than 1,000 units have been deployed to major research institutions worldwide ¹, used for VLA (Vision-Language-Action) model training, force control research, and human-robot interaction research.

Core Specifications Table

Item	Franka Research 3 (FR3)	Franka Panda (Legacy)
Manufacturer	Franka Robotics (under Agile Robots AG)	Franka Emika
Degrees of Freedom (DoF)	7-axis	7-axis
Payload	3 kg	3 kg
Reach	855 mm	850 mm
Repeatability	±0.1 mm (ISO 9283)	±0.1 mm
Workspace Coverage	94.5%	-
Maximum End-Effector Speed	2 m/s	2 m/s
Control Frequency	1 kHz	1 kHz
Joint Speed (J1-4)	150°/s	150°/s
Joint Speed (J5-7)	301°/s	301°/s
Robot Weight	17.8 kg	18 kg
Controller Weight	7 kg	-
Force Resolution	< 0.05 N	< 0.05 N
Protection Rating	IP40 (robot) / IP20 (controller)	IP30
Operating Temperature	+5°C ~ +45°C (optimal: 15-25°C)	+15°C ~ +25°C
Power	100-240V AC, 50-60Hz, 80W	-
Price	~$40,000+	~$30,000+ (discontinued)

Above specifications based on official datasheet and product page. ²

Joint Rotation Range

Joint	Minimum	Maximum
Joint 1	-166°	+166°
Joint 2	-101°	+101°
Joint 3	-166°	+166°
Joint 4	-176°	-4°
Joint 5	-166°	+166°
Joint 6	-1°	+215°
Joint 7	-166°	+166°

Joint ranges based on Franka Research 3 Datasheet v1.1. ³

Key Significance

1. Reference Platform for AI and Robotics Research

Franka Research 3 has been adopted as the de facto standard research platform by major universities and research institutes worldwide. World-leading institutions such as MIT, Stanford, CMU, and Oxford Robotics Institute use FR3 as core research equipment, forming the foundation for data collection for VLA model training.

2. Core Platform for VLA/Imitation Learning Research

Franka robot data constitutes a core portion of large-scale robot learning datasets including Open X-Embodiment and DROID. In recent research, VLA models using FR3 have achieved over 99% success rates on LIBERO benchmark and precision manipulation tasks (GPU insertion/removal) ⁴, demonstrating capability for uninterrupted operation for over 1 hour.

3. Popularization of Torque Control-Based Precision Force Control

With torque sensors embedded in all 7 joints, real-time external force estimation and collision detection are possible. This enables implementation of advanced force control techniques such as impedance control and compliance control, with intuitive hand guiding functionality also supported.

4. Research-Friendly Openness

1kHz real-time control is available through FCI (Franka Control Interface), supporting various development environments including libfranka C++ library, ROS 2, and MATLAB/Simulink. Researchers can freely implement from low-level control to high-level tasks.

Key Improvements Over Panda

Hardware Improvements

Item	Panda	FR3	Improvement
Reach	850 mm	855 mm	5mm increase
Workspace Coverage	Undisclosed	94.5%	Optimized kinematics
Weight	18 kg	17.8 kg	200g weight reduction
Joint Marking	None	X/Y axis markings	Intuitive setup
Design	-	Modern aesthetics	Ergonomic improvements
Protection Rating	IP30	IP40	Improved dust protection

Software Improvements

Desk API: REST interface for programmatic robot control (brake release, FCI activation, safety violation reset, etc.)
On-site Torque Sensor Recalibration: Researchers can calibrate torque sensors directly (maintains consistent sensitivity throughout lifecycle)
ROS 2 Jazzy Support: Full compatibility with latest ROS 2 framework
MoveIt 2 Integration: Full support for latest motion planning framework
OTA Updates: Over-the-air software update support

Control Interface (FCI)

Franka Control Interface (FCI) is FR3’s core low-level control interface, providing 1kHz real-time control through the libfranka C++ library.

Control Modes (5 types)

┌─────────────────────────────────────────────────────────────────┐
│          Franka Control Interface (FCI) Control Modes           │
├─────────────────────────────────────────────────────────────────┤
│  1. Joint Torque Control                                        │
│     └─ Direct joint torque command with gravity and friction    │
│        compensation                                             │
│                                                                 │
│  2. Joint Position Control                                      │
│     └─ Joint space position command                             │
│                                                                 │
│  3. Joint Velocity Control                                      │
│     └─ Direct joint velocity command                            │
│                                                                 │
│  4. Cartesian Pose Control                                      │
│     └─ Cartesian space position/orientation command             │
│                                                                 │
│  5. Cartesian Velocity Control                                  │
│     └─ Cartesian velocity command                               │
└─────────────────────────────────────────────────────────────────┘

1kHz Sensor Data Access

Real-time access to the following data:

Joint Data: Position (q), velocity (dq), link-side torque sensor signal (τ)
External Force Estimation: External torque and force estimates (τ_ext, F_ext)
Collision Information: Collision detection and contact information
Robot Model: Inertia matrix M(q), Coriolis term C(q,dq), gravity vector g(q), Jacobian J(q)

Asynchronous Joint Position Updates

The latest FCI supports asynchronous joint position updates, enabling more flexible and compliant control workflows.

Programming Interfaces (3 Levels)

FR3 provides three levels of programming interfaces depending on research purpose.

1. Desk (High-level)

Workflow-based Visual Programming Interface

Browser-based GUI with drag and drop programming
Quick prototyping without programming knowledge
Suitable for human-robot interaction research, demos, experiment setup
Extensible through Research Apps

2. RIDE (Mid-level)

Robot Integrated Development Environment

Custom Desk Apps development environment
External sensor and other feature connections
Lingua Franka: Domain-specific language (DSL) for state machine programming
ride-cli: Command line interface for development
Suitable for research not requiring millisecond-level reactions (task planning, path planning, perception-based manipulation)

3. FCI (Low-level)

Franka Control Interface - Real-time Low-level Control

Based on libfranka C++ library
1kHz real-time control command transmission
Essential for motion control research, advanced control algorithm development
Enables custom impedance control, force control implementation

Desk API (REST Interface)

Programmatic control added in 2025 update:

POST /api/brakes/release    # Release brakes
POST /api/fci/activate      # Activate FCI
POST /api/safety/clear      # Clear safety violations

Automated research workflow implementation
Enables unattended experiment environment setup
Documentation accessible at <robot IP>/deskapi

Safety Features

FR3 complies with ISO 10218 and ISO/TS 15066 safety standards as a collaborative robot.

Core Safety Mechanisms

Collision Detection
- Real-time external force detection based on 7-joint torque sensors
- Immediate collision response with fast control loop (1kHz)
- Software-based virtual walls/workspace limits
Power and Force Limiting
- Force/pressure limits based on ISO/TS 15066
- Custom safety rule configuration
- Self-damage prevention through internal monitoring
Safety Feature Configuration
- Custom safety rule implementation (worker protection)
- Speed, force, workspace limit programming
- Remote safety violation reset (Desk API)
Hand Guiding
- Intuitive manual manipulation based on torque sensors
- Smooth compliance control

VLA Research Applications

FR3 is one of the most widely used robot platforms in VLA (Vision-Language-Action) model research.

Major Research Results

Research	Achievement	FR3 Usage
Self-Improving VLA (2025)	Over 99% success rate, 1-hour uninterrupted GPU insertion/removal	Core experimental platform
PLD Post-Training	99% on LIBERO benchmark, 50%+ improvement on SimplerEnv	Real robot validation
CoT-VLA (2025)	46.7% relative performance improvement (53.7% → 78.8%)	Franka-Tabletop evaluation
ReBot (2025)	17% Octo, 20% OpenVLA performance improvement	Real-to-Sim-to-Real validation

Major Datasets

Dataset	Scale	FR3/Panda Usage
Open X-Embodiment	1M+ episodes	Core data source
DROID	76,000+ demos	Major collection platform
BridgeData V2	60,000+ trajectories	Partially included
LIBERO	130 tasks	PyBullet simulation
Franka Kitchen	Various tasks	dm_control based

Simulation Environments

MuJoCo: Official Franka support
Isaac Sim: NVIDIA Omniverse-based USD assets provided
LIBERO: PyBullet-based long-horizon manipulation benchmark
panda-gym: OpenAI Gym compatible RL environment
RLBench: CoppeliaSim-based 100+ tasks

Software Ecosystem

Official Support

Platform	Support Level	Description
libfranka	Official	C++ real-time control library
franka_ros2	Official	ROS 2 Control integration
ROS 2 Jazzy	Official (2025)	Full latest ROS 2 support
MoveIt 2	Official (2025)	Motion planning framework
MATLAB/Simulink	Official	Control system design
MuJoCo	Official	Physics simulation

Community Tools

Tool	Description
Franky	High-level Python/C++ control library
franka-interface	CMU iamlab’s libfranka wrapper
panda-gym	OpenAI Gym compatible reinforcement learning environment

Comparison with Other Robots

Item	FR3	UR5e	KUKA LWR	Kinova Gen3
DoF	7	6	7	7
Payload	3 kg	5 kg	7 kg	4 kg
Reach	855 mm	850 mm	800 mm	902 mm
Repeatability	±0.1 mm	±0.03 mm	±0.1 mm	±0.1 mm
Torque Sensors	All joints	None	All joints	None
Force Control	Native	External F/T required	Native	Limited
1kHz Control	Supported	Limited	Supported	Not supported
ROS 2 Support	Excellent	Good	Good	Good
VLA Ecosystem	Largest	Moderate	Moderate	Moderate
Price	~$40K	~$35K	~$100K	~$35K

FR3’s Differentiators

Native Torque Control: Precise force control with torque sensors embedded in all joints
VLA Research Ecosystem: Largest research community and datasets
Open API: 1kHz real-time control and full dynamics model access
Research-Friendly: Three-tier interfaces (Desk, RIDE, FCI) supporting various research levels

AI Companion (Optional)

Franka AI Companion is an edge computing solution that can be used with FR3.

NVIDIA GPU accelerated computation
Integration with 1kHz real-time control
AI/robotics research workflow acceleration
Experiment environment setup without separate workstation

Company History

Year	Event
2016	Franka Emika GmbH founded (Munich)
2017	German Future Prize winner, Panda release
2018	Time “50 Best Inventions” selection
2022	Franka Research 3 (FR3) release
2023.08	Preliminary insolvency filing due to shareholder dispute
2023.11	Acquired by Agile Robots AG (~32 million euros)
2025.02	FR3 1,000 units production achieved
2025	ROS 2 Jazzy, MoveIt 2, Desk API support