Master's thesis: Resilient Vision-Language-Action Models for Real-World Collaborative Robots

Background Vision-Language-Action (VLA) models represent a major step forward in robotics by translating visual observations and natural language instructions directly into physical motor control. However, deploying large multimodal models on physical robots presents critical challenges: foundation models require substantial compute, whereas offloading reasoning across the cloud-edge continuum introduces unpredictable network latency, buffer depletion, and physical safety risks. This Master's thesis project is hosted by RISE Research Institutes of Sweden in Kista and conducted in collaboration with KTH Royal Institute of Technology. The research is directly embedded within the Digital Futures research project EdgeWise (Secure and Low-Latency Vision-Language-Action Models for Adversarial Environments). EdgeWise investigates how embodied AI and VLA architectures can be deployed, executed, and secured for mission-critical robotic systems operating under constrained connectivity and latency-sensitive conditions. Building upon existing distributed multi-agent VLA prototypes developed at RISE, this project addresses key bottlenecks identified in distributed robotic control: mitigating latency-induced action-buffer starvation during asynchronous policy execution, integrating real-time safety supervision and partial failure recovery, and scaling distributed VLA pipelines from benchtop manipulators to real-world collaborative robots. The project offers flexible sub-tracks and can accommodate individual students or pairs of students collaborating on complementary aspects of the system. Description This master's thesis focuses on designing, implementing, and benchmarking resilient, distributed Vision-Language-Action (VLA) architectures for real-world collaborative and industrial manipulators. The core goal is to enable robust physical execution under real-world constraints such as network latency, bandwidth fluctuations, and unexpected environmental disturbances. The project addresses a fundamental trade-off in embodied AI: how to leverage large multimodal foundation models for high-level semantic reasoning and planning while maintaining deterministic, high-frequency, and safe real-time motor control at the physical edge. Students will work with modular multi-agent frameworks (orchestrating cloud planners, edge action policies, and supervisor agents) and investigate physical deployment on modern collaborative robotic platforms equipped with multi-camera setups and edge AI accelerators (such as NVIDIA Jetson). Key research questions include: How can distributed VLA architectures be scaled from lightweight benchtop manipulators to collaborative/industrial robots while preserving trajectory smoothness and closed-loop responsiveness? How do varying network conditions across the cloud-edge continuum impact asynchronous action chunking, and how can adaptive chunking or prediction buffering mitigate buffer starvation? How can an autonomous safety and supervisory agent be designed to monitor VLA actions in real time, detect anomalies or impending execution failures, and enforce safe overrides or semantic retry loops? The thesis scope is designed to be modular and can be pursued by individual students or split between two students working in close collaboration. Potential tracks include, for example: Investigating the computational, latency, and kinematic scalability of distributed VLA architectures when transitioning from lightweight desktop arms to industrial collaborative robots. Designing autonomous supervisory agents capable of monitoring execution integrity, preventing unsafe physical commands, and orchestrating partial task recovery Overcoming cloud-to-edge network latency and jitter to guarantee continuous, drift free real-time robotic motion Main Tasks The specific focus of the project will be tailored according to the student's profile and interests, covering key phases across software development, machine learning, and physical hardware evaluation: Review existing literature on Vision-Language-Action (VLA) models, imitation learning policies (such as ACT and Diffusion Policy), and distributed cloud-edge robotics architectures. Familiarise with the existing RISE multi-agent software framework, dataset collection tools, and inter-agent communication interfaces. Collect, synchronise, and preprocess multimodal demonstration datasets (integrating multi-view camera streams, robot proprioception, and natural language instructions) on physical robotic manipulators. Implement and optimise policy inference and control pipelines on edge computing platforms (such as NVIDIA Jetson) interfacing with robotic hardware. Design and benchmark communication layers across the cloud-edge continuum, evaluating data serialisation, network latency, and buffer management under variable connectivity. Develop and evaluate supervisory or verification mechanisms to ensure physical safety, detect operational failures, and support robust task recovery. Conduct empirical physical experiments to benchmark end-to-end latency, trajectory smoothness, and task completion rates across multi-step manipulation benchmarks. Document system architectures, benchmark datasets, and empirical findings in a comprehensive Master's thesis report conforming to university academic guidelines. Present and demonstrate the working system at RISE and the partner university. Qualifications This thesis project is suitable for one or two Master's students in Computer Science, Robotics, Machine Learning, Embedded Systems, or a related engineering discipline. Requirements: Strong programming skills in Python and practical familiarity with modern deep learning frameworks (such as PyTorch). Solid theoretical understanding and practical coursework in machine learning, computer vision, and/or robotics. Experience with Linux environments, Git version control, and containerised or modular software architectures. High motivation to conduct physical experiments on