NVIDIA Jetson edge computer
Runs computer vision, neural-network inference, grasp planning and high-level ROS 2 coordination.
Projects · Robotics and AI
Intelligent Robotic Hand
A modular robotic hand and forearm platform combining mechanical engineering, sensing, real-time control, computer vision and edge AI.
What the project is
The Intelligent Robotic Hand is a modular robotic manipulation platform intended to combine mechanical design, sensing, deterministic motor control, computer vision and edge artificial intelligence.
The central research question is whether a hand can adapt its grasp to objects with different shapes, orientations, weights and fragility without requiring an engineer to program every finger movement separately.
Development begins with measurable sub-systems, including human-hand motion capture, an index-finger and thumb mechanism, force and position sensing, and real-time motor control. These results will inform the design of the complete multi-finger prototype.
Development evidence
Hand sensing and motion-capture experiments
These images illustrate early work used to understand human-hand movement, joint relationships and sensor input before those behaviours are translated into a robotic mechanism.
Potential value
Why an intelligent robotic hand could be useful
Traditional grippers are excellent when an object and task are predictable. A sensor-rich adaptive hand may create value where products, positions or handling requirements change.
●Flexible handling of objects with different shapes and sizes
●Force-aware grasping for fragile or deformable items
●Packaging, sorting, assembly and logistics automation
●Research and education in robotics, control and machine learning
●A reusable end-effector platform for automation integrators
●Future assistive-technology research, subject to safety, clinical and regulatory validation
Technology stack
Hardware and software working as one control system
The architecture separates high-level perception and planning from time-critical local motor control. This allows advanced AI processing without sacrificing deterministic response and safety.
Hardware
ARM real-time controller
An STM32H7, Teensy-class or equivalent controller executes deterministic motor loops, sensor acquisition and local safety logic.
Actuators and motor drivers
The complete concept may use approximately fifteen actuators, with each joint requiring suitable drive electronics and current protection.
Position and force sensing
Encoders, Hall sensors or potentiometers provide position feedback, while current, pressure, force or tactile sensors support controlled contact.
External analogue acquisition
ADC hardware such as ADS1115, ADS1015 or MCP3008 can acquire analogue force, flex and pressure signals.
Industrial communication and safety
EtherCAT or another validated deterministic link transfers commands and feedback, supported by watchdogs, limits and emergency-stop behaviour.
Software
Linux, CUDA and TensorRT
The Jetson software stack accelerates computer vision and neural-network inference at the edge.
Python, C++ and OpenCV
Used for image processing, perception, data preparation, testing and application-level control.
PyTorch and model optimisation
Supports training or integrating perception and learning models, followed by optimisation for edge deployment.
ROS 2
Provides modular nodes for vision, grasp planning, hand control, telemetry, logging and system integration.
FreeRTOS or bare-metal firmware
Runs local motor-control tasks, sensor sampling, fault handling and watchdog logic on the ARM controller.
PID and adaptive control
Independent high-frequency control loops regulate joint position, velocity or force, with future scope for learning-based adaptation.
Control workflow
How the system is intended to operate
A grasp command becomes a coordinated sequence of perception, planning, deterministic control and sensor feedback.
Validation
How the prototype should be evaluated
Success should be demonstrated with measurable engineering evidence rather than attractive demonstrations alone, because robots have suffered enough from marketing departments.
Grasp success rate across a defined object test set
Position and force-control error
Maximum contact force on fragile objects
Response latency and complete grasp cycle time
Communication timing, jitter and packet-loss behaviour
Fault detection and recovery performance
Repeatability across repeated trials
Energy consumption, payload and actuator temperature
Related live workspace
Explore the RML Lab
View related research, modelling, simulation and engineering development in the live RML environment.
Development roadmap
A staged route from research to an industry pilot
Phase 1
Mechanism validation
Validate linkage, force, actuation, sensing, control and safety.
Phase 2
Integrated hand
Combine multiple digits, perception, planning and industrial communication.
Phase 3
Industry pilot
Test a narrow application and quantify performance, cost and value.