We are implementing differentiable models of robot manipulators, which allows us to learn typically assumed to be known models of robots for control and motion planning.

Implement Differentiable Kinematics Tree & Planning Objectives in PyTorch given URDF robot models.

One approach to explaining hierarchical levels of understanding within machine learning is the symbolic method of inductive logic programming (ILP), capable of learning logical rules that entail instances. ILP constructs logic programs from sets of ground atoms and background assumptions. Positive and negative instances are included which help define target predicates. The aim is to construct logic programs that explain provided positive sets and reject negative sets. Works have expanded ILP to allow end-to-end differentiable learning, resulting in research that combines deep learning architectures with relational models derived from ILP. However, as with the original framework, variants of differentiable ILP have focused primarily on learning discrete properties with little attention given to learning on continuous domains nor capturing non-linear properties. Given the importance of non-linearity, integrating such concepts is an open challenge still faced by ILP inspired frameworks. Capturing such non-linear properties is significantly important to a range of high impact scientific fields including robotics, physics, and computational biology, thus in this work we combine continuous and non-linear predicate functions with an extension to a differentiable ILP framework, so-called differentiable neural logic (dNL) networks. This is accomplished by exploring standard non-linear functions such as the power function and sine function, as well as operation functions such as multiplication and addition. Provided a tabular dataset comprised of non-linear mathematical relations, our non-linear dNL solver extracts the non-linear functions and assesses their performance on the training set to produce explanatory logical rules.

OpenCV (Computer vision) Model to differentiate poles, cones, robots, etc. so the robot doesn't crash into anything.

Final project for Robotic Science and Systems (CS 4610/5335). Using reinforcement learning to train a drone to use a birds-eye view camera to track and follow a person below. Target will be wearing a colored helmet to differentiate. The trained model will then be interfaced with a Duckie-Drone and track in a real-world setting.

documentation for differentiable-robot-model library

Differentiable episodic memory for long-horizon robot manipulation with vision-language-action models

🤖 Implement Actor-Critic Differentiable Model Predictive Control to enhance robot control using deep reinforcement learning and differentiable MPC techniques.

Code for "Differentiable Surrogate Models for Design and Trajectory Optimization of Auxetic Soft Robots"

FTC Challenge 2023-24 CenterStage code for Autonomous and Driver-control involving robot motion control and an image differentiating model

Recognizing and differentiating human vs robot in an online bidding war. Followed a complete Data Science lifecycle starting from EDA, Feature Engineering and Modelling

This project compares Random Forest and Logistic Regression classifiers and selects the best-performing model to predict and differentiate between healthy operation and faults (Robot Protective Stop) for a UR3 Cobot. The selected model achieves an F1-Score of 0.53 and a real-time inference latency of 6.47 ms.

Developed a system using machine learning algorithms to differentiate between a human and a robot based on 7.6 million bids made in various auctions. Implemented XGBoost and Random Forest machine learning algorithms on the attributed dataset and achieved a 94% efficiency using Python libraries such as keras, numpy, pandas and sklearn. Validated both models using k-fold cross validation on the training dataset achieving a 95% efficiency.