Open Calls

This project explores the use of Neural Architecture Search (NAS) and knowledge distillation to optimise deep learning models for perception tasks in lunar robotic missions. The goal is to design neural models for rock segmentation, obstacle detection, and terrain classification that automatically adapt and optimise for efficient execution on resource-constrained Edge AI devices (e.g., NVIDIA Jetson platforms). The work involves selecting representative lunar-analogue perception tasks, designing appropriate search spaces, and applying NAS and knowledge distillation methods to identify architectures that balance accuracy, robustness, and computational efficiency. The resulting models will undergo hardware-aware optimisation, including quantisation and pruning, to ensure real-time performance within the strict power and memory constraints typical of the computers on potential planetary rovers.

• Algorithm Development: Design NAS workflows tailored to lunar robotic perception tasks. Define search spaces and optimisation objectives for models such as rock segmentation, obstacle detection, and terrain classification. Implement and evaluate NAS strategies that effectively trade off accuracy, latency, and resource consumption.

• Hardware Optimisation and Benchmarks: Adapt the NAS-generated models for efficient deployment on Edge AI platforms (e.g., Jetson, MemryX, Axeler, Hailo). Apply hardware-aware model compression techniques, including pruning, quantisation, and TensorRT optimisation, to achieve real-time inference within tight computational and energy budgets.

• An algorithmic framework based on Neural Architecture Search for generating perception models optimised for lunar robotic applications and constrained onboard computing resources.

• A ready-to-use pipeline for hardware-aware model optimisation and deployment on Edge AI devices, producing efficient, real-time perception modules (e.g., rock segmentation) suitable for integration into lunar rover navigation systems.

This project aims to significantly expand the capabilities of an existing, UE5-based Lunar Environment Simulator to support advanced robotic perception and navigation research. The primary objective is to bridge the simulation-to-reality gap by integrating realistic rendering with a robust ROS2 interface and enabling the generation of comprehensive ground truth data, including depth maps and semantic segmentation labels, to facilitate the training of deep learning algorithms. The methodology involves designing simulation environments, ranging from procedurally generated challenging terrains to digital twins of local 3D-scanned analogue facilities. Furthermore, the work will investigate and attempt to implement real-time terramechanics models to simulate wheel-soil interactions, specifically addressing rover slip phenomena to improve the fidelity of virtual navigation experiments. The project builds on Spaceship EAC's prior experience with realistic rendering for space VR applications and on Spaceship Poland's previous experience building ROS 2 bridge interfaces for the LunarSim rover simulator.

• System Integration and Data Generation: Develop a seamless ROS 2 bridge to enable direct communication between the simulator and standard robotic control stacks. Implement advanced rendering pipelines to export synchronised multi-modal sensor data (RGB, Depth) alongside precise semantic ground truth annotations, enabling the rigorous training and validation of perception models.
• Environment Modelling and Physics: Create diverse and challenging lunar surface models, incorporating high-resolution 3D scans of physical testbeds to establish a "digital twin" environment. Research and integrate real-time terramechanics algorithms to simulate dynamic physical interactions, such as wheel slip and soil deformation, ensuring the simulated robot exhibits realistic kinematic behaviour.

• An enhanced simulation platform fully integrated with ROS 2, capable of generating photorealistic sensor streams and pixel-perfect semantic ground truth for machine learning applications and reference trajectories for navigation evaluation.
• A library of lunar environments, including a digital twin of an analogue site, coupled with a simple real-time terramechanics module that roughly replicates wheel-soil interactions for valid navigation testing.

This project investigates the use of general-purpose visual features extracted by deep neural networks, such as DINOv2, to estimate terrain traversability. The objective is to develop a method that leverages image-derived feature representations combined with a few-shot metric learning approach to generate traversability maps of lunar-like environments. The methodology involves creating sparse image annotations consisting of point pairs labelled according to their relative traversability (e.g., similar, better, or worse). These annotations support the optimisation of a metric-learning model that infers passability from visual features. The resulting model will be deployed on an Edge AI device to demonstrate real-time traversability estimation directly onboard a mobile robot. As a final component, the project will integrate the algorithm’s predicted traversability levels with depth-data processing to form a combined traversability assessment system.

• Algorithm development: Design and implement a traversability estimation method that leverages general visual features extracted by deep neural models. Develop a few-shot metric learning framework using sparse image annotations, constructed from point pairs with relative traversability labels, to generate reliable traversability predictions for lunar-like terrain.

• Hardware optimisation: Adapt and optimise the trained model for deployment on an Edge AI platform. This includes model compression, quantisation, or architectural adjustments to ensure real-time inference within the computational constraints of onboard hardware.

• Robot deployment and Testing: Integrate the optimised model into a mobile robot system and conduct field experiments. Validate the system’s performance by combining visual-based traversability estimations with depth-image data and assessing real-world passability mapping capabilities in lunar-analogue environments.

• An algorithm capable of estimating terrain traversability from general-purpose visual features using a lightweight few-shot metric learning approach.

• A ready-to-use pipeline that performs onboard, real-time traversability inference on an Edge AI device and integrates visual and depth-based information to produce a local passability map suitable for mobile robot navigation.

This project explores the use of Neural Architecture Search (NAS) and knowledge distillation to optimise deep learning models for perception tasks in lunar robotic missions. The goal is to design neural models for rock segmentation, obstacle detection, and terrain classification that automatically adapt and optimise for efficient execution on resource-constrained Edge AI devices (e.g., NVIDIA Jetson platforms). The work involves selecting representative lunar-analogue perception tasks, designing appropriate search spaces, and applying NAS and knowledge distillation methods to identify architectures that balance accuracy, robustness, and computational efficiency. The resulting models will undergo hardware-aware optimisation, including quantisation and pruning, to ensure real-time performance within the strict power and memory constraints typical of the computers on potential planetary rovers.

• Algorithm Development: Design NAS workflows tailored to lunar robotic perception tasks. Define search spaces and optimisation objectives for models such as rock segmentation, obstacle detection, and terrain classification. Implement and evaluate NAS strategies that effectively trade off accuracy, latency, and resource consumption.

• Hardware Optimisation and Benchmarks: Adapt the NAS-generated models for efficient deployment on Edge AI platforms (e.g., Jetson, MemryX, Axeler, Hailo). Apply hardware-aware model compression techniques, including pruning, quantisation, and TensorRT optimisation, to achieve real-time inference within tight computational and energy budgets.

• An algorithmic framework based on Neural Architecture Search for generating perception models optimised for lunar robotic applications and constrained onboard computing resources.

• A ready-to-use pipeline for hardware-aware model optimisation and deployment on Edge AI devices, producing efficient, real-time perception modules (e.g., rock segmentation) suitable for integration into lunar rover navigation systems.

This project aims to significantly expand the capabilities of an existing, UE5-based Lunar Environment Simulator to support advanced robotic perception and navigation research. The primary objective is to bridge the simulation-to-reality gap by integrating realistic rendering with a robust ROS2 interface and enabling the generation of comprehensive ground truth data, including depth maps and semantic segmentation labels, to facilitate the training of deep learning algorithms. The methodology involves designing simulation environments, ranging from procedurally generated challenging terrains to digital twins of local 3D-scanned analogue facilities. Furthermore, the work will investigate and attempt to implement real-time terramechanics models to simulate wheel-soil interactions, specifically addressing rover slip phenomena to improve the fidelity of virtual navigation experiments. The project builds on Spaceship EAC's prior experience with realistic rendering for space VR applications and on Spaceship Poland's previous experience building ROS 2 bridge interfaces for the LunarSim rover simulator.

• System Integration and Data Generation: Develop a seamless ROS 2 bridge to enable direct communication between the simulator and standard robotic control stacks. Implement advanced rendering pipelines to export synchronised multi-modal sensor data (RGB, Depth) alongside precise semantic ground truth annotations, enabling the rigorous training and validation of perception models.
• Environment Modelling and Physics: Create diverse and challenging lunar surface models, incorporating high-resolution 3D scans of physical testbeds to establish a "digital twin" environment. Research and integrate real-time terramechanics algorithms to simulate dynamic physical interactions, such as wheel slip and soil deformation, ensuring the simulated robot exhibits realistic kinematic behaviour.

• An enhanced simulation platform fully integrated with ROS 2, capable of generating photorealistic sensor streams and pixel-perfect semantic ground truth for machine learning applications and reference trajectories for navigation evaluation.
• A library of lunar environments, including a digital twin of an analogue site, coupled with a simple real-time terramechanics module that roughly replicates wheel-soil interactions for valid navigation testing.

This project investigates the use of general-purpose visual features extracted by deep neural networks, such as DINOv2, to estimate terrain traversability. The objective is to develop a method that leverages image-derived feature representations combined with a few-shot metric learning approach to generate traversability maps of lunar-like environments. The methodology involves creating sparse image annotations consisting of point pairs labelled according to their relative traversability (e.g., similar, better, or worse). These annotations support the optimisation of a metric-learning model that infers passability from visual features. The resulting model will be deployed on an Edge AI device to demonstrate real-time traversability estimation directly onboard a mobile robot. As a final component, the project will integrate the algorithm’s predicted traversability levels with depth-data processing to form a combined traversability assessment system.

• Algorithm development: Design and implement a traversability estimation method that leverages general visual features extracted by deep neural models. Develop a few-shot metric learning framework using sparse image annotations, constructed from point pairs with relative traversability labels, to generate reliable traversability predictions for lunar-like terrain.

• Hardware optimisation: Adapt and optimise the trained model for deployment on an Edge AI platform. This includes model compression, quantisation, or architectural adjustments to ensure real-time inference within the computational constraints of onboard hardware.

• Robot deployment and Testing: Integrate the optimised model into a mobile robot system and conduct field experiments. Validate the system’s performance by combining visual-based traversability estimations with depth-image data and assessing real-world passability mapping capabilities in lunar-analogue environments.

• An algorithm capable of estimating terrain traversability from general-purpose visual features using a lightweight few-shot metric learning approach.

• A ready-to-use pipeline that performs onboard, real-time traversability inference on an Edge AI device and integrates visual and depth-based information to produce a local passability map suitable for mobile robot navigation.