Aidan Scannell

Postdoctoral researcher | Machine learning | Sequential decision making | Robotics

Aidan Scannell

Postdoctoral Researcher

Biography

Hello, my name is Aidan Scannell and I am a postdoctoral researcher with interests at the intersection of machine learning, sequential decision-making, and robotics. My research focuses on developing autonomous agents capable of learning behaviors to solve a wide range of tasks. I am particularly interested in using natural language instructions to guide these agents and advancing robotics foundation models, especially foundation world models, to enable agents to solve new challenges quickly and effectively.

Bio:

I am a Finnish Center for Artificial Intelligence postdoctoral researcher at Aalto University in Joni Pajarinen’s Robot Learning Lab and Arno Solin’s Machine Learning Research Group. I obtained my PhD from the University of Bristol under the supervision of Arthur Richards and Carl Henrik Ek. During my PhD I developed methods for controlling quadcopters in uncertain environments by synergising methods from probabilistic machine learning, stochastic differential geometry and reinforcement learning.

Interests

Reinforcement learning
Embodied AI
Representation learning
World models
Robotics

Education

PhD Robotics and Autonomous Systems, 2022
University of Bristol, UK
MEng Mechanical Engineering, 2016
University of Bristol, UK

Recent News

[12.10.24] Giving a talk at Nordic AI Meet + AI Day 2024 - Sample-Efficient Reinforcement Learning with Implicitly Quantized Representations

[23.07.24] Giving a talk at IWIALS 2024 - iQRL: Implicitly Quantized Representations for Sample-Efficient Reinforcement Learning

[24.06.24] New paper accepted to ICML 2024 Workshop on Aligning Reinforcement Learning Experimentalists and Theorists (ARLET) - “Quantized Representations Prevent Dimensional Collapse in Self-predictive RL”

[19.06.24] Giving a lecture on “Model-based RL” at the Cambridge Ellis Unit Summer School on Probabilistic Machine Learning 2024

[12.06.24] New paper on arXiv - iQRL - Implicitly Quantized Representations for Sample-efficient Reinforcement Learning

ALL NEWS»

Recent Publications

Quickly discover relevant content by filtering publications.

Yi Zhao, Aidan Scannell, Tianyu Cui, Le Chen, Arno Solin, Juho Kannala, Joni Pajarinen (2024). Leveraging Unlabeled Offline Data via Model-based Active Semi-supervised Reinforcement Learning. Preprint.

Cite

Aidan Scannell, Kalle Kujanpää, Yi Zhao, Mohammadreza Nakhaei, Arno Solin, Joni Pajarinen (2024). iQRL - Implicitly Quantized Representations for Sample-efficient Reinforcement Learning. arXiv preprint arXiv:2406.02696.

PDF Cite Website

Aidan Scannell, Kalle Kujanpää, Yi Zhao, Mohammadreza Nakhaei, Arno Solin, Joni Pajarinen (2024). Quantized Representations Prevent Dimensional Collapse in Self-predictive RL. ICML Workshop on Aligning Reinforcement Learning Experimentalists and Theorists (ARLET).

Cite Website

Mohammadreza Nakhaei, Aidan Scannell, Joni Pajarinen (2024). Residual Learning and Context Encoding for Adaptive Offline-to-Online Reinforcement Learning. 4th Annual Conference on Learning for Dynamics and Control (L4DC).

PDF Cite Code

Aidan Scannell, Riccardo Mereu, Paul Chang, Ella Tamir, Joni Pajarinen, Arno Solin (2024). Function-space Parameterization of Neural Networks for Sequential Learning. The Twelth International Conference on Learning Representations (ICLR).

PDF Cite Code Poster Slides Website

See all publications

Projects

Aidan Scannell, Kalle Kujanpää, Yi Zhao, Mohammadreza Nakhaei, Arno Solin, Joni Pajarinen

Mar 24, 2024

Implicitly Quantized Representations for Reinforcement Learning

Learning representations for reinforcement learning (RL) has shown much promise for continuous control. In this project, we investigate using vector quantization to prevent representation collapse when learning representations for RL using a self-supervised latent-state consistency loss.

Aidan Scannell, Riccardo Mereu, Paul Chang, Ella Tamir, Joni Pajarinen, Arno Solin

Jan 10, 2024

Function-Space Bayesian Deep Learning for Sequential Learning

Sequential learning paradigms pose challenges for gradient-based deep learning due to difficulties incorporating new data and retaining prior knowledge. While Gaussian processes elegantly tackle these problems, they struggle with scalability and handling rich inputs, such as images.

Aidan Scannell, Arno Solin, Joni Pajarinen

Nov 19, 2022

Investigating Bayesian Neural Network Dynamics Models for Model-Based Reinforcement Learning

This project seeks to evaluate and compare different approaches for learning dynamics models in model-based RL. In particular, we plan to compare different approximate inference techniques (e.g. Laplace approximation, MC dropout, variational inference), as well as ensemble methods, to understand why they either succeed or fail in different environments.

Aidan Scannell, Carl Henrik Ek, Arthur Richards

Sep 24, 2022

Mode-Constrained Exploration for Model-Based Reinforcement Learning

This work presents a learning-based control method for navigating to a target state in unknown, or partially unknown, multimodal dynamical systems. In particular, it develops a model-based reinforcement learning algorithm that can remain in a desired dynamics mode with high probability. For example, if some of the dynamics modes are believed to be inoperable.

Aidan Scannell

May 6, 2021

GPJax - Gaussian Processes in Jax

Minimal Gaussian process library in JAX with a simple (custom) approach to state management.

Aidan Scannell

Nov 16, 2020

Trajectory Optimisation in Learned Multimodal Dynamical Systems

This work presents a two-stage method to perform trajectory optimisation in multimodal dynamical systems with unknown nonlinear stochastic transition dynamics. The method finds trajectories that remain in a preferred dynamics mode where possible and in regions of the transition dynamics model that have been observed and can be predicted confidently.

Aidan Scannell

Nov 16, 2020

Identifiable Mixtures of Sparse Variational Gaussian Process Experts

This work introduces a variational lower bound for the Mixture of Gaussian Process Experts model with a GP-based gating network based on sparse GPs. The model (and inference) are implemented in GPflow/TensorFlow.

Aidan Scannell

Mar 13, 2019

Probabilistic Modelling

I am in the process of creating Jupyter notebooks for several probabilistic models (Bayesian linear regression, Gaussian process regression) and approximate inference algorithms. Particular focus has been put on providing detailed theory as well as easy to follow code.

Aidan Scannell

Mar 13, 2019

Approximate Inference

This work implements and compares a variety of approximate inference techniques for the tasks of image de-noising (restoration) and image segmentation.

Aidan Scannell

Mar 13, 2019

Model-Based Reinforcement Learning with Gaussian Processes

In this work I re-implemented the PILCO algorithm in python using Tensorflow and GPflow. This work was mainly carried out for personal development and some of the implementation is based on this Python implementation. This repository will mainly serve as a baseline for my future research.

Aidan Scannell

Sep 20, 2018

Uncertain Agentspeak

During the first (taught) year of the FARSCOPE CDT program I conducted my masters thesis under the supervision of Professor Weiru Liu and Dr Kevin McAreavey titled “Extending BDI Agents to Model and Reason with Uncertainty”. I implemented and extended the AgentSpeak(L) (agent-based programing) language to enable agents to model and reason with uncertainty in a computationally efficient manner.

Aidan Scannell

Apr 26, 2018

Amazon Picking Challenge

As part of the FARSCOPE CDT program I worked in a team to develop a solution to Amazon’s picking challenge. This involved designing a robotic pick-and-place system that was capable of recognising and grasping both known and novel objects in cluttered environments.

Aidan Scannell

Jan 7, 2017

Autonomous Vehicle Lane Detection Software

An application was designed following the model-view-controller architecture to enable multiple autonomous vehicle algorithms to be simulated in different views and to allow the input parameters to be altered in run time e.g. adaptive threshold parameters, coordinates for inverse perspective mapping, number of sample points etc. The code will run slower due to the MVC architecture.

Aidan Scannell

Dec 29, 2016

Kidnapped Robot

This project involved developing algorithms capable of localising a robot within a known environment but at an unknown position and moving it to a target location. This was achieved in simulation using the BotSim library in Matlab and then implemented onto a real robot.

Aidan Scannell

Dec 29, 2016

Ultrasonic Non-Destructive Testing

This project entailed the design of an ultrasonic phased array for operation into the human body using Matlab.

Aidan Scannell

Nov 10, 2016

UAV Swarm

This project involved developing distributed software enabling a swarm of fixed wing UAVs to track a pollutant cloud. A discrete time state space model of the world was produced and a finite state machine was used to add intelligence.