Gerardo Aragón-Camarasa

Zijun Long, George Killick, Richard McCreadie et al.

Robotic vision applications often necessitate a wide range of visual perception tasks, such as object detection, segmentation, and identification. While there have been substantial advances in these individual tasks, integrating specialized models into a unified vision pipeline presents significant engineering challenges and costs. Recently, Multimodal Large Language Models (MLLMs) have emerged as novel backbones for various downstream tasks. We argue that leveraging the pre-training capabilities of MLLMs enables the creation of a simplified framework, thus mitigating the need for task-specific encoders. Specifically, the large-scale pretrained knowledge in MLLMs allows for easier fine-tuning to downstream robotic vision tasks and yields superior performance. We introduce the RoboLLM framework, equipped with a BEiT-3 backbone, to address all visual perception tasks in the ARMBench challenge-a large-scale robotic manipulation dataset about real-world warehouse scenarios. RoboLLM not only outperforms existing baselines but also substantially reduces the engineering burden associated with model selection and tuning. The source code is publicly available at https://github.com/longkukuhi/armbench.

Zijun Long, Zaiqiao Meng, Gerardo Aragon Camarasa et al.

Vision Transformers (ViTs) have emerged as popular models in computer vision, demonstrating state-of-the-art performance across various tasks. This success typically follows a two-stage strategy involving pre-training on large-scale datasets using self-supervised signals, such as masked random patches, followed by fine-tuning on task-specific labeled datasets with cross-entropy loss. However, this reliance on cross-entropy loss has been identified as a limiting factor in ViTs, affecting their generalization and transferability to downstream tasks. Addressing this critical challenge, we introduce a novel Label-aware Contrastive Training framework, LaCViT, which significantly enhances the quality of embeddings in ViTs. LaCViT not only addresses the limitations of cross-entropy loss but also facilitates more effective transfer learning across diverse image classification tasks. Our comprehensive experiments on eight standard image classification datasets reveal that LaCViT statistically significantly enhances the performance of three evaluated ViTs by up-to 10.78% under Top-1 Accuracy.

Zijun Long, George Killick, Lipeng Zhuang et al.

Image classification datasets exhibit a non-negligible fraction of mislabeled examples, often due to human error when one class superficially resembles another. This issue poses challenges in supervised contrastive learning (SCL), where the goal is to cluster together data points of the same class in the embedding space while distancing those of disparate classes. While such methods outperform those based on cross-entropy, they are not immune to labeling errors. However, while the detrimental effects of noisy labels in supervised learning are well-researched, their influence on SCL remains largely unexplored. Hence, we analyse the effect of label errors and examine how they disrupt the SCL algorithm's ability to distinguish between positive and negative sample pairs. Our analysis reveals that human labeling errors manifest as easy positive samples in around 99% of cases. We, therefore, propose D-SCL, a novel Debiased Supervised Contrastive Learning objective designed to mitigate the bias introduced by labeling errors. We demonstrate that D-SCL consistently outperforms state-of-the-art techniques for representation learning across diverse vision benchmarks, offering improved robustness to label errors.

Zijun Long, George Killick, Richard McCreadie et al.

State-of-the-art image models predominantly follow a two-stage strategy: pre-training on large datasets and fine-tuning with cross-entropy loss. Many studies have shown that using cross-entropy can result in sub-optimal generalisation and stability. While the supervised contrastive loss addresses some limitations of cross-entropy loss by focusing on intra-class similarities and inter-class differences, it neglects the importance of hard negative mining. We propose that models will benefit from performance improvement by weighting negative samples based on their dissimilarity to positive counterparts. In this paper, we introduce a new supervised contrastive learning objective, SCHaNe, which incorporates hard negative sampling during the fine-tuning phase. Without requiring specialized architectures, additional data, or extra computational resources, experimental results indicate that SCHaNe outperforms the strong baseline BEiT-3 in Top-1 accuracy across various benchmarks, with significant gains of up to $3.32\%$ in few-shot learning settings and $3.41\%$ in full dataset fine-tuning. Importantly, our proposed objective sets a new state-of-the-art for base models on ImageNet-1k, achieving an 86.14\% accuracy. Furthermore, we demonstrate that the proposed objective yields better embeddings and explains the improved effectiveness observed in our experiments.

Lipeng Zhuang, Shiyu Fan, Yingdong Ru et al.

We present Flat'n'Fold, a novel large-scale dataset for garment manipulation that addresses critical gaps in existing datasets. Comprising 1,212 human and 887 robot demonstrations of flattening and folding 44 unique garments across 8 categories, Flat'n'Fold surpasses prior datasets in size, scope, and diversity. Our dataset uniquely captures the entire manipulation process from crumpled to folded states, providing synchronized multi-view RGB-D images, point clouds, and action data, including hand or gripper positions and rotations. We quantify the dataset's diversity and complexity compared to existing benchmarks and show that our dataset features natural and diverse manipulations of real-world demonstrations of human and robot demonstrations in terms of visual and action information. To showcase Flat'n'Fold's utility, we establish new benchmarks for grasping point prediction and subtask decomposition. Our evaluation of state-of-the-art models on these tasks reveals significant room for improvement. This underscores Flat'n'Fold's potential to drive advances in robotic perception and manipulation of deformable objects. Our dataset can be downloaded at https://cvas-ug.github.io/flat-n-fold

Lipeng Zhuang, Shiyu Fan, Florent P. Audonnet et al.

We present Masked Generative Policy (MGP), a novel framework for visuomotor imitation learning. We represent actions as discrete tokens, and train a conditional masked transformer that generates tokens in parallel and then rapidly refines only low-confidence tokens. We further propose two new sampling paradigms: MGP-Short, which performs parallel masked generation with score-based refinement for Markovian tasks, and MGP-Long, which predicts full trajectories in a single pass and dynamically refines low-confidence action tokens based on new observations. With globally coherent prediction and robust adaptive execution capabilities, MGP-Long enables reliable control on complex and non-Markovian tasks that prior methods struggle with. Extensive evaluations on 150 robotic manipulation tasks spanning the Meta-World and LIBERO benchmarks show that MGP achieves both rapid inference and superior success rates compared to state-of-the-art diffusion and autoregressive policies. Specifically, MGP increases the average success rate by 9% across 150 tasks while cutting per-sequence inference time by up to 35x. It further improves the average success rate by 60% in dynamic and missing-observation environments, and solves two non-Markovian scenarios where other state-of-the-art methods fail.

Zijun Long, George Killick, Lipeng Zhuang et al.

State-of-the-art pre-trained image models predominantly adopt a two-stage approach: initial unsupervised pre-training on large-scale datasets followed by task-specific fine-tuning using Cross-Entropy loss~(CE). However, it has been demonstrated that CE can compromise model generalization and stability. While recent works employing contrastive learning address some of these limitations by enhancing the quality of embeddings and producing better decision boundaries, they often overlook the importance of hard negative mining and rely on resource intensive and slow training using large sample batches. To counter these issues, we introduce a novel approach named CLCE, which integrates Label-Aware Contrastive Learning with CE. Our approach not only maintains the strengths of both loss functions but also leverages hard negative mining in a synergistic way to enhance performance. Experimental results demonstrate that CLCE significantly outperforms CE in Top-1 accuracy across twelve benchmarks, achieving gains of up to 3.52% in few-shot learning scenarios and 3.41% in transfer learning settings with the BEiT-3 model. Importantly, our proposed CLCE approach effectively mitigates the dependency of contrastive learning on large batch sizes such as 4096 samples per batch, a limitation that has previously constrained the application of contrastive learning in budget-limited hardware environments.

Zijun Long, Kangheng Liang, Gerardo Aragon-Camarasa et al.

Interactive Text-to-image retrieval (I-TIR) is an important enabler for a wide range of state-of-the-art services in domains such as e-commerce and education. However, current methods rely on finetuned Multimodal Large Language Models (MLLMs), which are costly to train and update, and exhibit poor generalizability. This latter issue is of particular concern, as: 1) finetuning narrows the pretrained distribution of MLLMs, thereby reducing generalizability; and 2) I-TIR introduces increasing query diversity and complexity. As a result, I-TIR solutions are highly likely to encounter queries and images not well represented in any training dataset. To address this, we propose leveraging Diffusion Models (DMs) for text-to-image mapping, to avoid finetuning MLLMs while preserving robust performance on complex queries. Specifically, we introduce Diffusion Augmented Retrieval (DAR), a framework that generates multiple intermediate representations via LLM-based dialogue refinements and DMs, producing a richer depiction of the user's information needs. This augmented representation facilitates more accurate identification of semantically and visually related images. Extensive experiments on four benchmarks show that for simple queries, DAR achieves results on par with finetuned I-TIR models, yet without incurring their tuning overhead. Moreover, as queries become more complex through additional conversational turns, DAR surpasses finetuned I-TIR models by up to 7.61% in Hits@10 after ten turns, illustrating its improved generalization for more intricate queries.

Zijun Long, Lipeng Zhuang, George Killick et al.

Human-annotated vision datasets inevitably contain a fraction of human mislabelled examples. While the detrimental effects of such mislabelling on supervised learning are well-researched, their influence on Supervised Contrastive Learning (SCL) remains largely unexplored. In this paper, we show that human-labelling errors not only differ significantly from synthetic label errors, but also pose unique challenges in SCL, different to those in traditional supervised learning methods. Specifically, our results indicate they adversely impact the learning process in the ~99% of cases when they occur as false positive samples. Existing noise-mitigating methods primarily focus on synthetic label errors and tackle the unrealistic setting of very high synthetic noise rates (40-80%), but they often underperform on common image datasets due to overfitting. To address this issue, we introduce a novel SCL objective with robustness to human-labelling errors, SCL-RHE. SCL-RHE is designed to mitigate the effects of real-world mislabelled examples, typically characterized by much lower noise rates (<5%). We demonstrate that SCL-RHE consistently outperforms state-of-the-art representation learning and noise-mitigating methods across various vision benchmarks, by offering improved resilience against human-labelling errors.

Paul McHard, Florent P. Audonnet, Oliver Summerell et al.

Surface defects are a primary source of yield loss in manufacturing, yet existing anomaly detection methods often fail in real-world deployment due to limited and unrepresentative datasets. To overcome this, we introduce 3D-ADAM, a 3D Anomaly Detection in Additive Manufacturing dataset, that is the first large-scale, industry-relevant dataset for RGB+3D surface defect detection in additive manufacturing. 3D-ADAM comprises 14,120 high-resolution scans of 217 unique parts, captured with four industrial depth sensors, and includes 27,346 annotated defects across 12 categories along with 27,346 annotations of machine element features in 16 classes. 3D-ADAM is captured in a real industrial environment and as such reflects real production conditions, including variations in part placement, sensor positioning, lighting, and partial occlusion. Benchmarking state-of-the-art models demonstrates that 3D-ADAM presents substantial challenges beyond existing datasets. Validation through expert labelling surveys with industry partners further confirms its industrial relevance. By providing this benchmark, 3D-ADAM establishes a foundation for advancing robust 3D anomaly detection capable of meeting manufacturing demands.

Oliver Summerell, Gerardo Aragon-Camarasa, Stephanie Ordonez Sanchez

This paper investigates the capability of Neural Networks (NNs) as alternatives to the traditional methods to analyse the performance of aerofoils used in the wind and tidal energy industry. The current methods used to assess the characteristic lift and drag coefficients include Computational Fluid Dynamics (CFD), thin aerofoil and panel methods, all face trade-offs between computational speed and the accuracy of the results and as such NNs have been investigated as an alternative with the aim that it would perform both quickly and accurately. As such, this paper provides a benchmark for the windAI_bench dataset published by the National Renewable Energy Laboratory (NREL) in the USA. In order to validate the methodology of the benchmarking, the AirfRANSdataset benchmark is used as both a starting point and a point of comparison. This study evaluates four neural networks (MLP, PointNet, GraphSAGE, GUNet) trained on a range of aerofoils at 25 angles of attack (4$^\circ$ to 20$^\circ$) to predict fluid flow and calculate lift coefficients ($C_L$) via the panel method. GraphSAGE and GUNet performed well during the training phase, but underperformed during testing. Accordingly, this paper has identified PointNet and MLP as the two strongest models tested, however whilst the results from MLP are more commonly correct for predicting the behaviour of the fluid, the results from PointNet provide the more accurate results for calculating $C_L$.

Zijun Long, George Killick, Richard McCreadie et al.

As Multimodal Large Language Models (MLLMs) grow in size, adapting them to specialized tasks becomes increasingly challenging due to high computational and memory demands. Indeed, traditional fine-tuning methods are costly, due to the need for extensive, task-specific training. While efficient adaptation methods exist that aim to reduce these costs, in practice they suffer from shallow inter-modal alignment, which severely hurts model effectiveness. To tackle these computational challenges and improve inter-modal alignment, we introduce the MultiWay-Adapter (MWA), a novel framework featuring an 'Alignment Enhancer'. This enhancer deepens inter-modal alignment, enabling high transferability with minimal tuning effort. Our experiments show that unlike prior efficient tuning approaches, MWA maintains model effectiveness, while reducing training time by up-to 57%. MWA is also lightweight, increasing model size by only 2-3% (in terms of parameters) for state-of-the-art foundation models like BEiT-3 Large. These results demonstrate that MWA provides an efficient and effective adaptation method for MLLMs, significantly broadening their applicability.

Li Duan, Lewis Boyd, Gerardo Aragon-Camarasa

In this paper, we propose to predict the physics parameters of real fabrics and garments by learning their physics similarities between simulated fabrics via a Physics Similarity Network (PhySNet). For this, we estimate wind speeds generated by an electric fan and the area weight to predict bending stiffness of simulated and real fabrics and garments. We found that PhySNet coupled with a Bayesian optimiser can predict physics parameters and improve the state-of-art by 34%for real fabrics and 68% for real garments.

Li Duan, Gerardo Aragon-Camarasa

We present a Garment Similarity Network (GarNet) that learns geometric and physical similarities between known garments by continuously observing a garment while a robot picks it up from a table. The aim is to capture and encode geometric and physical characteristics of a garment into a manifold where a decision can be carried out, such as predicting the garment's shape class and its visually perceived weight. Our approach features an early stop strategy, which means that GarNet does not need to observe a garment being picked up from a crumpled to a hanging state to make a prediction. In our experiments, we find that GarNet achieves prediction accuracies of 92% for shape classification and 95.5% for predicting weights and advances state-of-art approaches by 21% for shape classification.

Luca Marzari, Ameya Pore, Diego Dall'Alba et al.

Deep Reinforcement Learning (DRL) is emerging as a promising approach to generate adaptive behaviors for robotic platforms. However, a major drawback of using DRL is the data-hungry training regime that requires millions of trial and error attempts, which is impractical when running experiments on robotic systems. Learning from Demonstrations (LfD) has been introduced to solve this issue by cloning the behavior of expert demonstrations. However, LfD requires a large number of demonstrations that are difficult to be acquired since dedicated complex setups are required. To overcome these limitations, we propose a multi-subtask reinforcement learning methodology where complex pick and place tasks can be decomposed into low-level subtasks. These subtasks are parametrized as expert networks and learned via DRL methods. Trained subtasks are then combined by a high-level choreographer to accomplish the intended pick and place task considering different initial configurations. As a testbed, we use a pick and place robotic simulator to demonstrate our methodology and show that our method outperforms a benchmark methodology based on LfD in terms of sample-efficiency. We transfer the learned policy to the real robotic system and demonstrate robust grasping using various geometric-shaped objects.

Ali AlQallaf, Gerardo Aragon-Camarasa

While humans are aware of their body and capabilities, robots are not. To address this, we present in this paper a neural network architecture that enables a dual-arm robot to get a sense of itself in an environment. Our approach is inspired by human self-awareness developmental levels and serves as the underlying building block for a robot to achieve awareness of itself while carrying out tasks in an environment. We assume that a robot has to know itself before interacting with the environment in order to be able to support different robotic tasks. Hence, we implemented a neural network architecture to enable a robot to differentiate its limbs from the environment using visual and proprioception sensory inputs. We demonstrate experimentally that a robot can distinguish itself with an accuracy of 88.7% on average in cluttered environmental settings and under confounding input signals.

Li Duan, Gerardo Aragon-Camarasa

We present an approach to continuous perception for robotic laundry tasks. Our assumption is that the visual prediction of a garment's shapes and weights is possible via a neural network that learns the dynamic changes of garments from video sequences. Continuous perception is leveraged during training by inputting consecutive frames, of which the network learns how a garment deforms. To evaluate our hypothesis, we captured a dataset of 40K RGB and 40K depth video sequences while a garment is being manipulated. We also conducted ablation studies to understand whether the neural network learns the physical and dynamic properties of garments. Our findings suggest that a modified AlexNet-LSTM architecture has the best classification performance for the garment's shape and weights. To further provide evidence that continuous perception facilitates the prediction of the garment's shapes and weights, we evaluated our network on unseen video sequences and computed the 'Moving Average' over a sequence of predictions. We found that our network has a classification accuracy of 48% and 60% for shapes and weights of garments, respectively.

Nikos Pitsillos, Ameya Pore, Bjorn Sand Jensen et al.

We present an introspective framework inspired by the process of how humans perform introspection. Our working assumption is that neural network activations encode information, and building internal states from these activations can improve the performance of an actor-critic model. We perform experiments where we first train a Variational Autoencoder model to reconstruct the activations of a feature extraction network and use the latent space to improve the performance of an actor-critic when deciding which low-level robotic behaviour to execute. We show that internal states reduce the number of episodes needed by about 1300 episodes while training an actor-critic, denoting faster convergence to get a high success value while completing a robotic task.

Ameya Pore, Gerardo Aragon-Camarasa

We present a behaviour-based reinforcement learning approach, inspired by Brook's subsumption architecture, in which simple fully connected networks are trained as reactive behaviours. Our working assumption is that a pick and place robotic task can be simplified by leveraging domain knowledge of a robotics developer to decompose and train such reactive behaviours; namely, approach, grasp, and retract. Then the robot autonomously learns how to combine them via an Actor-Critic architecture. The Actor-Critic policy is to determine the activation and inhibition mechanisms of the reactive behaviours in a particular temporal sequence. We validate our approach in a simulated robot environment where the task is picking a block and taking it to a target position while orienting the gripper from a top grasp. The latter represents an extra degree-of-freedom of which current end-to-end reinforcement learning fail to generalise. Our findings suggest that robotic learning can be more effective if each behaviour is learnt in isolation and then combined them to accomplish the task. That is, our approach learns the pick and place task in 8,000 episodes, which represents a drastic reduction in the number of training episodes required by an end-to-end approach and the existing state-of-the-art algorithms.

Li Sun, Gerardo Aragon-Camarasa, Simon Rogers et al.

This paper proposes a single-shot approach for recognising clothing categories from 2.5D features. We propose two visual features, BSP (B-Spline Patch) and TSD (Topology Spatial Distances) for this task. The local BSP features are encoded by LLC (Locality-constrained Linear Coding) and fused with three different global features. Our visual feature is robust to deformable shapes and our approach is able to recognise the category of unknown clothing in unconstrained and random configurations. We integrated the category recognition pipeline with a stereo vision system, clothing instance detection, and dual-arm manipulators to achieve an autonomous sorting system. To verify the performance of our proposed method, we build a high-resolution RGBD clothing dataset of 50 clothing items of 5 categories sampled in random configurations (a total of 2,100 clothing samples). Experimental results show that our approach is able to reach 83.2\% accuracy while classifying clothing items which were previously unseen during training. This advances beyond the previous state-of-the-art by 36.2\%. Finally, we evaluate the proposed approach in an autonomous robot sorting system, in which the robot recognises a clothing item from an unconstrained pile, grasps it, and sorts it into a box according to its category. Our proposed sorting system achieves reasonable sorting success rates with single-shot perception.

Li Sun, Gerardo Aragon-Camarasa, Simon Rogers et al.

This paper presents a novel robot vision architecture for perceiving generic 3D clothes configurations. Our architecture is hierarchically structured, starting from low-level curvatures, across mid-level geometric shapes \& topology descriptions; and finally approaching high-level semantic surface structure descriptions. We demonstrate our robot vision architecture in a customised dual-arm industrial robot with our self-designed, off-the-self stereo vision system, carrying out autonomous grasping and dual-arm flattening. It is worth noting that the proposed dual-arm flattening approach is unique among the state-of-the-art robot autonomous system, which is the major contribution of this paper. The experimental results show that the proposed dual-arm flattening using stereo vision system remarkably outperforms the single-arm flattening and widely-cited Kinect-based sensing system for dexterous manipulation tasks. In addition, the proposed grasping approach achieves satisfactory performance on grasping various kind of garments, verifying the capability of proposed visual perception architecture to be adapted to more than one clothing manipulation tasks.

Gerardo Aragon-Camarasa, Susanne B. Oehler, Yuan Liu et al.

To provide insight into cloth perception and manipulation with an active binocular robotic vision system, we compiled a database of 80 stereo-pair colour images with corresponding horizontal and vertical disparity maps and mask annotations, for 3D garment point cloud rendering has been created and released. The stereo-image garment database is part of research conducted under the EU-FP7 Clothes Perception and Manipulation (CloPeMa) project and belongs to a wider database collection released through CloPeMa (www.clopema.eu). This database is based on 16 different off-the-shelve garments. Each garment has been imaged in five different pose configurations on the project's binocular robot head. A full copy of the database is made available for scientific research only at https://sites.google.com/site/ugstereodatabase/.