RP1 Introduction

RP1 is RoboParty's fully open-source humanoid robot platform. From mechanical structures, circuit boards, and wiring harnesses to firmware, device drivers, robot models, and training and deployment code, every layer of RP1 is fully open-source and published in the GitHub repository. You can study a single subsystem, build a robot from the published materials, or use the platform to explore your own control methods and learning workflows.

We hope to make robots trusted companions to humanity.

RP1 is built for individual developers, research teams, and education. We aim to provide a high-performance robot that makes algorithm research easier, and a more open robot platform that makes teaching easier. RP1 aims to provide that shared starting point.

RP1 is open across three layers:

The interfaces between these layers are open as well. The robot model used for training, the exported policy, and the physical robot's motor mapping share the same joint definitions, with the mapping at each stage recorded in configuration files. A change at any layer can be traced through this chain to the others.

The RP1 series offers two configurations: RP1 and RP1 Pro. The technical specifications below match the RP1 product page, with standard and optional equipment identified separately.

• Performance and operating time vary with load, terrain, battery condition, and software configuration.
• Heads, end effectors, and perception sensors vary by model and selected options. Refer to the selected configuration for included equipment.

Two open-source models are available, distinguished by their body motion joints; head and finger joints are excluded. These 24 / 28-axis counts refer to the body only, unlike the active degrees of freedom listed above: RP1 Pro has 28 body joints and 2 head joints for a total of 30 active degrees of freedom. Dexterous-hand finger degrees of freedom are counted separately.

Compared with the 24-axis model, the 28-axis model adds Yaw and Pitch joints at each wrist. Both models are provided in URDF and MJCF formats and can be loaded directly into MuJoCo or Isaac Lab.

Note The 24-axis configuration currently provides a complete path from training through simulation transfer checks to deployment on the physical robot. The 28-axis deployment configuration is ready, but its default policy files have not yet been released, and no dedicated training tasks are available yet.

Modular. The head and hand units can be changed to suit the task, while the legs and waist provide a common locomotion base. End effectors can use RoboParty's own RP Hand dexterous hand or other actuators and sensors.

Symmetric and standardized. The upper limbs are fully symmetric, and the same lower-limb molds are used on both sides. Fewer spare-part types make maintenance, replacement, and upgrades easier.

The structure is the enclosure. The main structure also serves as the enclosure, reducing additional covers and weight. Openings can serve as carrying handles or temporary mounting points for cameras and other accessories during experiments. Internal airflow channels help dissipate heat.

Designed for human contact. Rounded contours are paired with a soft rubber coating, and the head is shaped for an easy grip. The blue-and-white color scheme avoids both the imposing feel of industrial equipment and the childish look of a toy, giving it a character closer to a household appliance.

Version 1.3 improvements. The hip Pitch motors have been moved upward and the legs brought closer together for more human-like proportions. The hand motors now use a serial arrangement for more compact end effectors, and the rotating joint cross-sections have changed from square to circular.

RP1's electrical system comprises six parts: power, main control, expansion, interaction, perception, and actuation. They connect through CAN-FD buses, USB 3.0, Gigabit Ethernet, and power buses.

• Dual power supplies. A 48V lithium battery pack is the main power source. It slides into the rear of the torso and directly supplies both the main control board and the rear interface board. A smaller 30V backup battery supports hot swapping: the rear interface board charges it during normal operation, and it supplies power in the reverse direction during a momentary loss of main power to briefly maintain critical control circuits. The main battery is charged using a 140W charging management board and a custom charging dock.
• Separate application computing and real-time control. The RBE-MAIN-RK3588S main control board has two processors. The RK3588 application processor (6 TOPS NPU, 8GB LPDDR4x, 32GB eMMC) handles inference, vision, speech, and interaction; a 480MHz real-time MCU handles motion scheduling, power sequencing, and hardware emergency stops. Neither preempts the other.
• Distributed joint control. Each joint has a built-in driver that closes the current, velocity, and position loops locally. The main controller sends target positions or torques over CAN-FD and receives position, temperature, and fault codes.
• An independent emergency-stop path. Under the architecture design, emergency-stop signals reach the real-time MCU over a hardwired connection or 433MHz link and directly disable the motors without passing through the operating system.
• Expandable computing. RP1 includes RK3588 ×1 as its base computing module, with an optional development module. RP1 Pro includes RK3588 ×1 + A733 and an Orin NX (16 GB) development module. The rear interface board reserves space for an NVIDIA Jetson NX expansion module and provides USB 3.0, Gigabit Ethernet, Type-C, and multiple power outputs.

Gerber files, BOMs, component placement coordinates, and 3D models for all 14 circuit boards are available in RP1_hardware. For a system overview, see the RP1 Hardware System Architecture PDF in the repository.

RP1's software has four layers, from bottom to top:

The drivers and robot interface are implemented in C++ with Python bindings (imu_py, motors_py, and robot_py). You can bypass the inference node and use a few lines of Python to read the IMU, control a single motor, or send actions to the whole robot.

The default deployment configuration uses a 4 ms host control step (250 Hz) and runs the policy every 5 steps (50 Hz). A single observation frame for the default 24-axis policy has 81 dimensions:

The policy receives a 243-dimensional vector formed by concatenating the latest 3 frames and outputs 24 joint target offsets. These are multiplied by an action scale of 0.25 and added to the default pose, then tracked by the joint drivers using PD control.

roboparty_train is a training workspace based on Isaac Sim 5.1.0, Isaac Lab 2.3.2, and RSL-RL. It currently registers the following RP1 tasks:

After training, play.py exports the policy in both TorchScript and ONNX formats. The TorchScript version can be used directly in MuJoCo for Sim2Sim checks, while the ONNX version is used for deployment on the physical robot. The deployment repository already includes 24-axis velocity-tracking and dance-motion-tracking policies, which you can use to learn the deployment workflow when you receive the robot.

RP Hand dexterous hand. A compact dexterous hand for small and medium-sized humanoid robots, with 6 actively controlled axes and 5 additional passive or coupled degrees of freedom, for a total of 11 mechanical degrees of freedom. It uses worm gearing, measures 147.54 × 73.53 × 26.35 mm, weighs 379 g ± 10 g, and offers optional tactile sensing. It connects through CAN-FD and comes with C++ / Python SDKs and ROS 2 nodes.

Gamepad and mobile app. The deployment software supports direct gamepad control: enable motors, reset the pose, start or stop inference, and switch motion sequences, with joysticks controlling walking direction and turning. The roboparty_joystick Android client sends control commands at 50 Hz over UDP or Bluetooth Low Energy. It includes virtual joysticks, an emergency-stop button with confirmation, and skills such as waving, bowing, and dancing, alongside live battery, IMU, and motor status.

Whole-robot testing platform. rp1_test publishes CAD and software for three types of load endurance test rigs covering the arms, torso, and legs, plus drafts of 18 whole-robot reliability test methods to assess whether an RP1 can operate reliably over extended periods.

Third-party files retain their original licenses. Refer to LICENSE and LICENSES.md in each repository for details.

Issues and pull requests are welcome in the relevant repository. Including the part number, board revision, or software version helps us identify problems more quickly.