Overview

The TriFinger is a robotic platform intended to support research in dexterous manipulation. All hardware and software (including a simulator) are open-source, links can be found below. Please check the paper for a detailed description of the hardware and software: https://arxiv.org/abs/2008.03596

Hardware

Each finger has 3 DoF and they share a workspace, permitting complex fine-manipulation. Three RGB cameras ensure good visibility for any configuration. Instructions for building your own platform can be found here.

The internal mechanics are based on the quadruped proposed here. This design has the following qualities:

• low weight, high torque

• 1 kHz torque control and sensing

• robustness to impacts due to transparency of transmission

Software

The key strengths of the software framework are:

• simple user interface in Python and C++ for control at up to 1 kHz

• safety checks to prevent the robot from breaking

• synchronized history of all inputs and outputs available and can be logged

• we provide a simulator of the proposed platform

The robot-agnostic code can be found in the robot_interfaces repository, see here for a demo of usage in C++, and here for a demo of how a new robot can be implemented. The drivers for our particular robot are implemented in robot_fingers. More detailed instructions for installing and using the code can be found in the documentation:

• Documentation of robot_interfaces

• Documentation of robot_fingers

For a demo of usage in Python, check this file, here is a snippet of actual code:

Simulator

We provide a simulator (based on PyBullet) of the TriFinger robot. The simulator provides an interface which is identical to the one of the real robot, which makes switching easy. Please see here for documentation and installation instructions.

Illustrative Experiments

To illustrate the capabilities of the platform, we perform simple demonstration, optimal control and deep reinforcement learning experiments.

Fine Manipulation

These motions were recorded through kinesthetic teaching (i.e. the motion was demonstrated by guiding the robot fingers).

Throwing

As above, these motions were recorded through kinesthetic teaching.

Optimal Control

Here, we execute a real-time 1kHz control loop which computes the optimal forces to be applied to the object.

Deep Reinforcement Learning

Here we apply an out-of-the-box implementation of a deep RL algorithm (DDPG from stable baselines) to learn reaching from scratch. Notably, no safety precautions are necessary on the user side.