Chalmers Open Digital Repository

Välkommen till Chalmers öppna digitala arkiv!

Här hittar du:

  • Studentarbeten utgivna på lärosätet, såväl kandidatarbeten som examensarbeten på grund- och masternivå
  • Digitala specialsamlingar, som t ex Chalmers modellkammare
  • Utvalda projektrapporter

Forskningspublikationer, rapporter och avhandlingar hittar du i research.chalmers.se

Enheter i Chalmers ODR

Välj en enhet för att se alla samlingar.

Senast publicerade

  • Bild (thumbnail)
    Exploring the Benefits of Compression to Improve Data-intensive Neural Network Inference
    (2026) Meidinger, Max Andreas Peter
    Convolutional Neural Networks (CNNs) have become essential for many computer vision tasks such as image classification, object detection and semantic segmentation. State-of-the-art models come with high computational demands that require efficient acceleration. FPGAs are a popular choice, as they can host energy-efficient, high-throughput inference designs. Streaming dataflow designs are particularly efficient, as they create custom per-layer pipelines which confine most of the data movement on chip. As these accelerators also prefer to keep model parameters on chip, on-chip buffer space becomes a limiting factor. Long skip connections, a ubiquitous structure of today’s networks, aggravate this problem further, as they require large amounts of buffer space to synchronize the diverging branches of the model’s architecture. Off-chip eviction instead moves the activations of skip connections temporarily to off-chip memory, but this puts additional load on the memory bus, which can lead to stalls in the compute pipeline if the processing engines cannot be fed. This problem can be alleviated using lossless compression, but lossless compression is fundamentally limited. This thesis explores lossy compression of the evicted activations instead, trading a bounded loss of inference accuracy for a higher compression ratio, without retraining. Starting from SZ1, a prediction-based compressor whose model fits the streaming paradigm, we develop ASZ (Activation SZ) in a design space exploration, which adds block-level error thresholds, new prediction functions, per-channel selection of the prediction functions, and Huffman coding and downsampling stages. We evaluate ASZ against lossless (Huffman coding, RLE) and lossy baselines (quantization with Huffman coding (QH), Lorenzo prediction, SZ1, downsampling, SVD and JPEG) on the 74 skip connections of eight CNN classifiers on ImageNet, CIFAR-10 and CIFAR-100, and estimate the hardware cost of every method. Lossy compression goes well beyond lossless encoding: on every connection, a lossy method exceeds the best lossless one, on 71 of them at a cost of at most 0 1 pp of top-1 accuracy. Within a budget of 1 pp, the streaming-compatible compressors QH, Lorenzo and ASZ reach geometric-mean compression ratios of 3 9, 3 8 and 3 5, 2 3 to 2 6 times the 1 5 of Huffman coding, while RLE does not compress on average. ASZ improves on SZ1 (2 7), but trails the quantizers on most connections, since SZ1-type compressors spend their bits on the choice of a prediction function rather than on the values or their residuals. JPEG reaches 6 6, but buffers eight rows per channel and needs multipliers on the data path. When several connections are compressed, their errors add up and the ratio within the budget falls, least for the error-bounded Lorenzo and SZ1; with three connections, the best lossy method still stays 1 5 to 4 3 times above Huffman coding. Lossy compression can thus substantially reduce the off-chip traffic of evicted skip connections, and QH, the cheapest lossy compressor and the best streaming-compatible one in ratio, offers the best trade-off between compression ratio and hardware cost.
  • Bild (thumbnail)
    Walking the Talk: Vision-Language Orchestration of Navigation and Motion-Imitation Skills on a Quadruped Robot
    (2026) Chen, Xiaoyang; Liao, Zifeng
    Campus guide robots are expected not only to guide visitors through complex environments but also to interact naturally, respond to changing intentions, and engage through expressive physical behaviour. Realising these capabilities requires tight coordination of conversational intelligence, navigation, and learned motor control. This thesis presents FIDO (Friendly and Interactive robot DOg), a quadruped campus guide built on the Unitree Go2 platform, integrating full-duplex conversation, vision-language orchestration, indoor–outdoor navigation, and learned motor skills within a unified perception–action framework. We develop ShallWeTalk for predictive turn-taking and interruption handling, and a Vision-Language Orchestrator that translates speech, visual observations, conversational context, and robot state into spoken responses and executable actions. A Robot Execution Backend executes actions and reports outcomes, while LiDAR localisation and RTK-GNSS/odometry fusion support indoor–outdoor navigation. To expand FIDO’s motor repertoire, we construct MyGO2, a library of 1,731 Go2 motion clips, and develop a general motion-tracking controller. We introduce two hand-designed reference motions: Crawl, which uses a lower body posture than any motion in MyGO2, and Moonwalk, an expressive backward-sliding movement. For Moonwalk, we train dedicated tracking policies using two approaches: asymmetric actor–critic learning and teacher–student distillation. We also develop EasyGO2, a common deployment framework supporting MuJoCo sim-to-sim validation and hardware execution. During live English conversations, FIDO stops speaking for eight of nine interruptions but continues through six of eight backchannels, tolerating pauses of up to 1.5 s without added response latency relative to a silence-based baseline. The general controller completes 163 of 176 unseen clips in simulation and executes Crawl zero-shot on the Go2; both Moonwalk approaches yield policies that succeed in all five hardware trials without hardware fine-tuning. FIDO relocalises in all 30 indoor trials and reaches all 20 navigation goals. Experts rate output quality of the Vision-Language Orchestrator at 98.5/100 across 149 tasks. Integrated campus demonstrations show FIDO following a 231 m outdoor route in response to a spoken request and, in a separate indoor run, changing its goal, pausing, and resuming on voice commands. Together, these results demonstrate how coordinated dialogue, navigation, and learned motor control enable natural, responsive, and physically expressive visitor interaction.
  • Bild (thumbnail)
    Design and Empirical Evaluation of a Multi-Agent System for Engineering Workflows
    (2026) Qiu, Mengbing; Zhang, Lingyue
    Early-phase truck integration requires engineers to coordinate vehicle data, regional regulations, engineering calculations, and CAD-related operations across heterogeneous tools. This thesis designs, implements, and empirically evaluates a toolorchestrated multi-agent system for a bounded set of these tasks. The research examines heterogeneous-resource integration through a traceable and extensible toolorchestrated architecture, correct workflow execution and contract conformance on the defined benchmark, and trade-offs in accuracy, evidence faithfulness, reliability, and latency across tasks and configurations. The prototype uses a LangGraph supervisor–specialist architecture comprising one orchestrator and five ReAct subagents for Database, Regulation, Payload, Agility, and CAD tasks. A separate non-agent utility node manages the RAG knowledge base. The study follows an artifact-centred design-science strategy. Evaluation is conducted at the offline, component, system, and implementation levels using frozen datasets. These datasets include a simulated vehicle database, a curated local regulation corpus, deterministic reference calculations, and explicit workflow checks. On the defined benchmark, routing accuracy was 91.0% across 100 requests, and 49 of 50 end-to-end tasks passed the required sequence and output checks. The Payload calculation results agreed with an independent oracle in all 60 evaluated cases. In the Database evaluation, the required values were retrieved more often than the generated result exactly matched the reference result. This difference demonstrates why capability-specific metrics are needed. At retrieval depth five, the annotated answer fact was found for 83.65% of regulation queries, whereas the specifically annotated source was retrieved for 64.42%. The paired automatic comparison between rule-only and RAG-only regulation answering did not detect a statistically significant difference in term match (p = 0.1263), while rule-only execution had a substantial latency advantage. The results support the feasibility of bounded multi-agent orchestration for engineering assistance. They do not establish production performance or superiority over a monolithic agent. The quantitative findings are limited to the synthetic benchmark and small document corpus. The CAD and Agility measurements cover planning rather than complete geometry execution. The study therefore supports a capability-specific interpretation in which deterministic operations, retrieved evidence, execution traces, and continued human oversight remain necessary for dependable engineering use
  • Bild (thumbnail)
  • Bild (thumbnail)
    Intelligent Sound Diagnostics - Acoustic Fault Detection in Truck Steering Systems Using Machine Learning and Deep Learning
    (2026) Krishnamoorthy, Kiruthiga
    This thesis investigates acoustic fault diagnosis for heavy duty truck steering components, focusing on two safety critical parts, the Power Steering Pump (PSP) and the Steering Gear, using a real world dataset of 326 recordings collected under operational conditions. The problem is formulated as a seven class classification task covering four sound types under normal and faulty conditions. Classical machine learning models trained on hand crafted acoustic features are compared against a convolutional neural network trained from scratch and Pretrained Audio Neural Network (PANN) models used as frozen feature extractors, to identify which modelling approach is most effective for this task. Seven individual acoustic descriptors, MFCC, log-Mel spectrogram statistics, sub-band energy ratios, spectral shape descriptors, time-domain statistics, amplitude envelope kurtosis, and discrete wavelet transform coefficients, are also evaluated in isolation to determine which is most discriminative for fault detection. All models are evaluated using StratifiedGroupKFold cross validation with file level grouping to prevent leakage between overlapping audio segments, and performance is reported using the Matthews Correlation Coefficient (MCC) and macro F1, both robust to class imbalance. A linear Support Vector Machine trained on a 180-feature ANOVA selected set achieves the strongest overall performance, reaching a held out test MCC of 1.000, ahead of every deep learning approach evaluated, including a pretrained PANN backbone with over 80 million parameters. A compact CNN trained entirely from scratch matches the best pretrained PANN variant despite requiring no external pretraining, while the largest pretrained backbone performs weakest of all models compared. Among the individual descriptors, log-Mel spectrogram statistics and MFCCs prove the most discriminative, while amplitude envelope kurtosis alone carries almost no discriminative power. These findings suggest that carefully engineered acoustic features combined with a simple classical classifier can outperform learned and transferred deep representations on a dataset of this scale, offering a lightweight and interpretable alternative for industrial predictive maintenance.