Constraint-Aware Zero-Shot Vision-Language Navigation in Continuous Environments

We address the task of Vision-Language Navigation in Continuous Environments (VLN-CE) under the zero-shot setting. Zero-shot VLN-CE is particularly challenging due to the absence of expert demonstrations for training and minimal environment structural prior to guide navigation. To confront these challenges, we propose a Constraint-Aware Navigator (CA-Nav), which reframes zero-shot VLN-CE as a sequential, constraint-aware sub-instruction completion process. CA-Nav continuously translates sub-instructions into navigation plans using two core modules: the Constraint-Aware Sub-instruction Manager (CSM) and the Constraint-Aware Value Mapper (CVM). CSM defines the completion criteria for decomposed sub-instructions as constraints and tracks navigation progress by switching sub-instructions in a constraint-aware manner. CVM, guided by CSM's constraints, generates a value map on the fly and refines it using superpixel clustering to improve navigation stability. CA-Nav achieves the state-of-the-art performance on two VLN-CE benchmarks, surpassing the previous best method by 12 percent and 13 percent in Success Rate on the validation unseen splits of R2R-CE and RxR-CE, respectively. Moreover, CA-Nav demonstrates its effectiveness in real-world robot deployments across various indoor scenes and instructions.

PhysGame: Uncovering Physical Commonsense Violations in Gameplay Videos

Recent advancements in video-based large language models (Video LLMs) have witnessed the emergence of diverse capabilities to reason and interpret dynamic visual content. Among them, gameplay videos stand out as a distinctive data source, often containing glitches that defy physics commonsense. This characteristic renders them an effective benchmark for assessing the under-explored capability of physical commonsense understanding in video LLMs. In this paper, we propose PhysGame as a pioneering benchmark to evaluate physical commonsense violations in gameplay videos. PhysGame comprises 880 videos associated with glitches spanning four fundamental domains (i.e., mechanics, kinematics, optics, and material properties) and across 12 distinct physical commonsense. Through extensively evaluating various state-ofthe-art video LLMs, our findings reveal that the performance of current open-source video LLMs significantly lags behind that of proprietary counterparts. To bridge this gap, we curate an instruction tuning dataset PhysInstruct with 140,057 question-answering pairs to facilitate physical commonsense learning. In addition, we also propose a preference optimization dataset PhysDPO with 34,358 training pairs, where the dis-preferred responses are generated conditioned on misleading titles (i.e., meta information hacking), fewer frames (i.e., temporal hacking) and lower spatial resolutions (i.e., spatial hacking). Based on the suite of datasets, we propose PhysVLM as a physical knowledge-enhanced video LLM. Extensive experiments on both physical-oriented benchmark PhysGame and general video understanding benchmarks demonstrate the state-ofthe-art performance of PhysVLM.

Rethinking Weight-Averaged Model-merging

Weight-averaged model-merging has emerged as a powerful approach in deep learning, capable of enhancing model performance without fine-tuning or retraining. However, the underlying mechanisms that explain its effectiveness remain largely unexplored. In this paper, we investigate this technique from three novel perspectives to provide deeper insights into how and why weight-averaged model-merging works: (1) we examine the intrinsic patterns captured by the learning of the model weights, through the visualizations of their patterns on several datasets, showing that these weights often encode structured and interpretable patterns; (2) we investigate model ensemble merging strategies based on averaging on weights versus averaging on features, providing detailed analyses across diverse architectures and datasets; and (3) we explore the impact on model-merging prediction stability in terms of changing the parameter magnitude, revealing insights into the way of weight averaging works as regularization by showing the robustness across different parameter scales. Our findings shed light on the "black box" of weight-averaged model-merging, offering valuable insights and practical recommendations that advance the model-merging process.

Continual LLaVA: Continual Instruction Tuning in Large Vision-Language Models

Instruction tuning constitutes a prevalent technique for tailoring Large Vision Language Models (LVLMs) to meet individual task requirements. To date, most of the existing approaches are confined to single-task adaptation, whereas the requirements in real-world scenarios are inherently varied and continually evolving. Thus an ideal LVLM should sustain continual instruction tuning in the face of stream-task distributions (i.e., different domains, emerging capabilities, and new datasets) while minimizing the forgetting of previously acquired knowledge. To achieve this, we propose a new benchmark for COntinuAl inStruction Tuning on LVLMs (COAST), which encompasses the aforementioned domain-incremental, capability-incremental, and dataset-incremental configurations. In terms of methodology, we propose Continual LLaVA, a rehearsal-free method tailored for continual instruction tuning in LVLMs. To circumvent the additional overhead associated with experience replay, we freeze LVLMs and construct the dual increment embeddings for each input instruction to facilitate parameter-efficient tuning. Specifically, the increment embeddings can be decomposed into two principal components: 1) intrinsic increment embeddings to encode task-specific characteristics. To achieve this, we set up a low-rank pool containing candidate embeddings, from which we select the relevant ones based on their similarity with the user instructions; 2) contextual increment embeddings to investigate the inter-dependencies across tasks. In this regard, the low-rank embeddings chosen in the previous tasks are aggregated via learnable weighted sum to provide complementary hints. Extensive experiments indicate that the proposed Continual LLaVA outperforms previous methods by significantly reducing the forgetting during the continual instruction tuning process.

BEVPose: Unveiling Scene Semantics through Pose-Guided Multi-Modal BEV Alignment

In the field of autonomous driving and mobile robotics, there has been a significant shift in the methods used to create Bird's Eye View (BEV) representations. This shift is characterised by using transformers and learning to fuse measurements from disparate vision sensors, mainly lidar and cameras, into a 2D planar ground-based representation. However, these learning-based methods for creating such maps often rely heavily on extensive annotated data, presenting notable challenges, particularly in diverse or non-urban environments where large-scale datasets are scarce. In this work, we present BEVPose, a framework that integrates BEV representations from camera and lidar data, using sensor pose as a guiding supervisory signal. This method notably reduces the dependence on costly annotated data. By leveraging pose information, we align and fuse multi-modal sensory inputs, facilitating the learning of latent BEV embeddings that capture both geometric and semantic aspects of the environment. Our pretraining approach demonstrates promising performance in BEV map segmentation tasks, outperforming fully-supervised state-of-the-art methods, while necessitating only a minimal amount of annotated data. This development not only confronts the challenge of data efficiency in BEV representation learning but also broadens the potential for such techniques in a variety of domains, including off-road and indoor environments.

Learn 2 Rage: Experiencing The Emotional Roller Coaster That Is Reinforcement Learning

This work presents the experiments and solution outline for our teams winning submission in the Learn To Race Autonomous Racing Virtual Challenge 2022 hosted by AIcrowd. The objective of the Learn-to-Race competition is to push the boundary of autonomous technology, with a focus on achieving the safety benefits of autonomous driving. In the description the competition is framed as a reinforcement learning (RL) challenge. We focused our initial efforts on implementation of Soft Actor Critic (SAC) variants. Our goal was to learn non-trivial control of the race car exclusively from visual and geometric features, directly mapping pixels to control actions. We made suitable modifications to the default reward policy aiming to promote smooth steering and acceleration control. The framework for the competition provided real time simulation, meaning a single episode (learning experience) is measured in minutes. Instead of pursuing parallelisation of episodes we opted to explore a more traditional approach in which the visual perception was processed (via learned operators) and fed into rule-based controllers. Such a system, while not as academically "attractive" as a pixels-to-actions approach, results in a system that requires less training, is more explainable, generalises better and is easily tuned and ultimately out-performed all other agents in the competition by a large margin.

Simultaneous Diffusion Sampling for Conditional LiDAR Generation

By enabling capturing of 3D point clouds that reflect the geometry of the immediate environment, LiDAR has emerged as a primary sensor for autonomous systems. If a LiDAR scan is too sparse, occluded by obstacles, or too small in range, enhancing the point cloud scan by while respecting the geometry of the scene is useful for downstream tasks. Motivated by the explosive growth of interest in generative methods in vision, conditional LiDAR generation is starting to take off. This paper proposes a novel simultaneous diffusion sampling methodology to generate point clouds conditioned on the 3D structure of the scene as seen from multiple views. The key idea is to impose multi-view geometric constraints on the generation process, exploiting mutual information for enhanced results. Our method begins by recasting the input scan to multiple new viewpoints around the scan, thus creating multiple synthetic LiDAR scans. Then, the synthetic and input LiDAR scans simultaneously undergo conditional generation according to our methodology. Results show that our method can produce accurate and geometrically consistent enhancements to point cloud scans, allowing it to outperform existing methods by a large margin in a variety of benchmarks.

Affordance-Centric Policy Learning: Sample Efficient and Generalisable Robot Policy Learning using Affordance-Centric Task Frames

Affordances are central to robotic manipulation, where most tasks can be simplified to interactions with task-specific regions on objects. By focusing on these key regions, we can abstract away task-irrelevant information, simplifying the learning process, and enhancing generalisation. In this paper, we propose an affordance-centric policy-learning approach that centres and appropriately \textit{orients} a \textit{task frame} on these affordance regions allowing us to achieve both \textbf{intra-category invariance} -- where policies can generalise across different instances within the same object category -- and \textbf{spatial invariance} -- which enables consistent performance regardless of object placement in the environment. We propose a method to leverage existing generalist large vision models to extract and track these affordance frames, and demonstrate that our approach can learn manipulation tasks using behaviour cloning from as little as 10 demonstrations, with equivalent generalisation to an image-based policy trained on 305 demonstrations. We provide video demonstrations on our project site: https://affordance-policy.github.io.

GraspMamba: A Mamba-based Language-driven Grasp Detection Framework with Hierarchical Feature Learning

Grasp detection is a fundamental robotic task critical to the success of many industrial applications. However, current language-driven models for this task often struggle with cluttered images, lengthy textual descriptions, or slow inference speed. We introduce GraspMamba, a new language-driven grasp detection method that employs hierarchical feature fusion with Mamba vision to tackle these challenges. By leveraging rich visual features of the Mamba-based backbone alongside textual information, our approach effectively enhances the fusion of multimodal features. GraspMamba represents the first Mamba-based grasp detection model to extract vision and language features at multiple scales, delivering robust performance and rapid inference time. Intensive experiments show that GraspMamba outperforms recent methods by a clear margin. We validate our approach through real-world robotic experiments, highlighting its fast inference speed.

Oops, I Sampled it Again: Reinterpreting Confidence Intervals in Few-Shot Learning

The predominant method for computing confidence intervals (CI) in few-shot learning (FSL) is based on sampling the tasks with replacement, i.e.\ allowing the same samples to appear in multiple tasks. This makes the CI misleading in that it takes into account the randomness of the sampler but not the data itself. To quantify the extent of this problem, we conduct a comparative analysis between CIs computed with and without replacement. These reveal a notable underestimation by the predominant method. This observation calls for a reevaluation of how we interpret confidence intervals and the resulting conclusions in FSL comparative studies. Our research demonstrates that the use of paired tests can partially address this issue. Additionally, we explore methods to further reduce the (size of the) CI by strategically sampling tasks of a specific size. We also introduce a new optimized benchmark, which can be accessed at https://github.com/RafLaf/FSL-benchmark-again