ByteMini: 22-DoF Bi-Manual Robotic Platform

• ByteMini is a purpose-built 22-DoF bi-manual robotic platform designed for dexterous, mobile manipulation with integrated advanced sensing.
• It features dual 7-DoF manipulators with quasi-direct-drive actuators and a mobile base, enabling precise and robust performance in diverse environments.
• The system leverages the GR-3 Vision-Language-Action model to seamlessly combine autonomous planning and teleoperation, achieving high success rates on unseen tasks.

ByteMini is a purpose-built, 22-degree-of-freedom (DoF) bi-manual mobile robotic platform designed as the primary real-world hardware for the GR-3 Vision-Language-Action (VLA) model. Developed to support generalist robot policies, ByteMini delivers flexible, reliable, and dexterous manipulation, enabling seamless integration with large-scale VLA models for complex, long-horizon tasks spanning generalization to unseen objects, environments, and instructions. Its development emphasizes mechanical robustness, precise actuation, comprehensive sensing, whole-body compliance control, and interoperability with advanced vision-LLMs to serve as an exemplar of a generalist robotic embodiment (Cheang et al., 21 Jul 2025).

1. Mechanical Structure and Actuation

ByteMini employs an architecture centered around versatility and dexterity, leveraging a high-DoF configuration and advanced materials:

• Degrees of Freedom and Configuration: Each arm consists of 7 revolute joints (7-DoF) arranged in a human-like kinematic chain with a compact “sphere” wrist. Both arms together yield a 14-DoF dual-manipulator, while the mobile base provides 3-DoF (translation in x, y, and rotation about yaw). Vertical lift and head camera assembly contribute an additional 5-DoF, totaling 22-DoF.
• Materials and Joints: All arm links are machined from 7075-T6 aluminum, selected for a high stiffness-to-weight ratio. The Gosselin–Lavoie sphere wrist design minimizes the volume near the end-effector and avoids singularities, enabling reliable function in constrained spaces.
• Actuation: Quasi-Direct-Drive (QDD) actuators in all arm joints supply low friction, high transparency, and intrinsic back-drivability, supporting both compliant control and responsive manipulation.
• End-Effector: Each arm terminates in an underactuated two-finger pinch gripper, supporting a 2 kg payload per gripper and featuring elastomer fingertip pads for conformal contact.
• Physical Metrics: Each arm may reach approximately 750 mm from the shoulder in the fully extended pose; the mobile base footprint is 600 mm × 550 mm. The chassis has a low center of mass for stability and an overall weight of roughly 65 kg, with a variable height range of 550–900 mm via the lift mechanism.

2. Kinematic and Dynamic Modeling

The kinematic and dynamic models underpin precise trajectory generation and real-time feedback control:

• Kinematic Description: Each 7-DoF manipulator is characterized via Denavit–Hartenberg (DH) parameters (table below). The last three $\alpha_i$ parameters of the sphere wrist are offset so all wrist axes intersect, maximizing orientational dexterity.

| i | $a_i$ (mm) | $\alpha_i$ (rad) | $d_i$ (mm) | $\theta_i$ | |---|------------|------------------|------------|------------------| | 1 | 0 | $+\pi/2$ | 330 | $q_1$ | | 2 | 350 | 0 | 0 | $q_2$ | | 3 | 0 | $+\pi/2$ | 0 | $q_3$ | | 4 | 0 | $a_i$0 | 315 | $a_i$1 | | 5 | 0 | $a_i$2 | 0 | $a_i$3 | | 6 | 0 | $a_i$4 | 80 | $a_i$5 | | 7 | 0 | 0 | 0 | $a_i$6 |

• Dynamic Model: The full 22-DoF system is described by

$a_i$7

where $a_i$8 represents all joint angles, $a_i$9 is the symmetric positive-definite mass/inertia matrix, $\alpha_i$0 comprises Coriolis and centrifugal terms, $\alpha_i$1 is gravity, and $\alpha_i$2 is the joint torque vector.

• Control Framework: Whole-body compliance control, following the method of Sentis & Khatib (2006), utilizes a quadratic program (QP) to enforce manipulability, singularity avoidance, joint-limit compliance, and external force regulation in real time. For teleoperation, human limb kinematics (via Meta VR Quest) are mapped to the robot’s workspace.

3. Sensor Suite and Perceptual Integration

ByteMini is equipped with a comprehensive multimodal sensor suite emphasizing high-bandwidth robotic perception:

• Visual Sensing: The head houses a single RGBD camera (640×480, 90° FOV) for scene-wide perception, while both wrists carry RGBD sensors of identical specification, positioned ~150 mm from the gripper and angled ±20° downward for close-range, low-latency feedback on hand-object interactions.
• Inertial and Force Sensing: An IMU on the mobile base enables drift-free odometry; joint torques are estimated from QDD actuator currents. There are no dedicated force/torque sensors at the wrists.
• Sensor Fusion: Multi-view fusion leverages the VLM backbone (Qwen2.5-VL-3B-Instruct), producing a unified representation across viewpoints—a bird’s-eye and hand-eye point cloud embedding—enabling the GR-3 model to act on rich, egocentric perceptual input.

4. Hardware Compute and Software Architecture

The ByteMini control stack merges on-board computation, real-time actuation, and a modular software ecosystem:

• Compute Hardware: An Intel NUC 11 (8-core CPU, 16 GB RAM) provides central compute, complemented by an integrated Iris Xe GPU for light image preprocessing. An EtherCAT real-time bus synchronously controls all 22 QDD motors. Dual-lithium batteries furnish 10 hours of operational runtime.
• Software Stack:
  • Middleware: ROS 2 Foxy serves as the middleware framework.
  • Motion Planning: MoveIt! 2 is employed for inverse-kinematics and fallback motion planning.
  • Policy Interface: The GR-3 inference server, implemented in C++, is exposed as a ROS 2 action server.
  • Trajectory Optimization: A custom C++ library solves a QP at 100 Hz to refine joint trajectories.
  • Teleoperation: Unity-based VR client communicates via a ROS 2 WebSocket bridge for immersive operator control.

5. Integration with GR-3 VLA Model

The synergy between ByteMini and GR-3 supports both autonomous and teleoperated learning, facilitating rapid adaptation:

• Control Loop: At each timestep $\alpha_i$3, the GR-3 model outputs an action chunk $\alpha_i$4 (joint-space waypoints). The onboard optimizer interpolates using a minimum-jerk QP and dispatches resulting torque commands $\alpha_i$5 via (1). For mobile base navigation, pure pursuit is utilized.
• Data Collection:
  • Robot Trajectories: Whole-body compliance controller enables teleoperation, collecting ~250 robot trajectories per hour.
  • Human Demonstrations: Human VR data (PICO 4 Ultra) delivers ~450 trajectories per hour, mapped to joint-space via inverse kinematics and view-padding.
• Fine-tuning Protocol: GR-3 undergoes few-shot fine-tuning on both robot and human demonstration data, in conjunction with web-scale VL data, enhancing sample efficiency, cross-embodiment generalization, and rapid task adaptation.

6. Task Performance and Experimental Results

Empirical evaluation establishes ByteMini’s capabilities across diverse manipulation tasks in real-world settings:

• Generalizable Pick-and-Place:
  • Seen objects: 99.3% success (Basic), 98.1% (Unseen Environment).
  • Unseen instructions: Success increases from 40% (baseline $\alpha_i$6) to 77.1% (GR-3).
  • Unseen objects: 57.8% to 86.7% success after 10-shot VR fine-tuning.
• Long-Horizon Table Bussing:
  • Flat setting (“clean up table”): 90.2% task progress (GR-3) vs. 85.4% (baseline).
  • Instruction-Following (sub-tasks): 97.5% success (GR-3) vs. 53.8% (baseline).
• Dexterous Cloth Hanging:
  • Basic: 86.7% task progress.
  • Challenging positions: 83.9%.
  • Unseen garments (short sleeves): 75.8%.

ByteMini, when operated under the GR-3 VLA policy, demonstrates robust, long-horizon, bi-manual, and mobile manipulation with strict language-conditioned instruction following and generalization to novel objects and environments. Its combination of mechanical design, sensor fusion, and compliance-based control provides the infrastructure required for advanced generalist robotic policies (Cheang et al., 21 Jul 2025).