Sample-efficient robot learning is a longstanding goal in robotics. Inspired by the success of scaling in vision and language, the robotics community is now investigating large-scale offline datasets for robot learning. However, existing methods often require expert and/or reward-labeled task-specific data, which can be costly and limit their application in practice. In this paper, we consider a more realistic setting where the offline data consists of reward-free and non-expert multi-embodiment offline data. We show that generalist world model pre-training (WPT), together with retrieval-based experience rehearsal and execution guidance, enables efficient reinforcement learning (RL) and fast task adaptation with such non-curated data. In experiments over 72 visuomotor tasks, spanning 6 different embodiments, covering hard exploration, complex dynamics, and various visual properties, WPT achieves 35.65% and 35% higher aggregated score compared to widely used learning-from-scratch baselines, respectively.

In real-world scenarios, achieving domain adaptation and generalization poses significant challenges, as models must adapt to or generalize across unknown target distributions. Extending these capabilities to unseen multimodal distributions, i.e., multimodal domain adaptation and generalization, is even more challenging due to the distinct characteristics of different modalities. Significant progress has been made over the years, with applications ranging from action recognition to semantic segmentation. Besides, the recent advent of large-scale pre-trained multimodal foundation models, such as CLIP, has inspired works leveraging these models to enhance adaptation and generalization performances or adapting them to downstream tasks. This survey provides the first comprehensive review of recent advances from traditional approaches to foundation models, covering: (1) Multimodal domain adaptation; (2) Multimodal test-time adaptation; (3) Multimodal domain generalization; (4) Domain adaptation and generalization with the help of multimodal foundation models; and (5) Adaptation of multimodal foundation models. For each topic, we formally define the problem and thoroughly review existing methods. Additionally, we analyze relevant datasets and applications, highlighting open challenges and potential future research directions. We maintain an active repository that contains up-to-date literature at https://github.com/donghao51/Awesome-Multimodal-Adaptation.

Gaussian splatting enables fast novel view synthesis in static 3D environments. However, reconstructing real-world environments remains challenging as distractors or occluders break the multi-view consistency assumption required for accurate 3D reconstruction. Most existing methods rely on external semantic information from pre-trained models, introducing additional computational overhead as pre-processing steps or during optimization. In this work, we propose a novel method, DeSplat, that directly separates distractors and static scene elements purely based on volume rendering of Gaussian primitives. We initialize Gaussians within each camera view for reconstructing the view-specific distractors to separately model the static 3D scene and distractors in the alpha compositing stages. DeSplat yields an explicit scene separation of static elements and distractors, achieving comparable results to prior distractor-free approaches without sacrificing rendering speed. We demonstrate DeSplat's effectiveness on three benchmark data sets for distractor-free novel view synthesis. See the project website at https://aaltoml.github.io/desplat/.

Object-Centric Learning (OCL) can discover objects in images or videos by simply reconstructing the input. For better object discovery, representative OCL methods reconstruct the input as its Variational Autoencoder (VAE) intermediate representation, which suppresses pixel noises and promotes object separability by discretizing continuous super-pixels with template features. However, treating features as units overlooks their composing attributes, thus impeding model generalization; indexing features with scalar numbers loses attribute-level similarities and differences, thus hindering model convergence. We propose \textit{Grouped Discrete Representation} (GDR) for OCL. We decompose features into combinatorial attributes via organized channel grouping, and compose these attributes into discrete representation via tuple indexes. Experiments show that our GDR improves both Transformer- and Diffusion-based OCL methods consistently on various datasets. Visualizations show that our GDR captures better object separability.

\textit{Relative overgeneralization} (RO) occurs in cooperative multi-agent learning tasks when agents converge towards a suboptimal joint policy due to overfitting to suboptimal behavior of other agents. In early work, optimism has been shown to mitigate the \textit{RO} problem when using tabular Q-learning. However, with function approximation optimism can amplify overestimation and thus fail on complex tasks. On the other hand, recent deep multi-agent policy gradient (MAPG) methods have succeeded in many complex tasks but may fail with severe \textit{RO}. We propose a general, yet simple, framework to enable optimistic updates in MAPG methods and alleviate the RO problem. Specifically, we employ a \textit{Leaky ReLU} function where a single hyperparameter selects the degree of optimism to reshape the advantages when updating the policy. Intuitively, our method remains optimistic toward individual actions with lower returns which are potentially caused by other agents' sub-optimal behavior during learning. The optimism prevents the individual agents from quickly converging to a local optimum. We also provide a formal analysis from an operator view to understand the proposed advantage transformation. In extensive evaluations on diverse sets of tasks, including illustrative matrix games, complex \textit{Multi-agent MuJoCo} and \textit{Overcooked} benchmarks, the proposed method\footnote{Code can be found at \url{https://github.com/wenshuaizhao/optimappo}.} outperforms strong baselines on 13 out of 19 tested tasks and matches the performance on the rest.

We present, QP-SBGD, a novel layer-wise stochastic optimiser tailored towards training neural networks with binary weights, known as binary neural networks (BNNs), on quantum hardware. BNNs reduce the computational requirements and energy consumption of deep learning models with minimal loss in accuracy. However, training them in practice remains to be an open challenge. Most known BNN-optimisers either rely on projected updates or binarise weights post-training. Instead, QP-SBGD approximately maps the gradient onto binary variables, by solving a quadratic constrained binary optimisation. Under practically reasonable assumptions, we show that this update rule converges with a rate of $\mathcal{O}(1 / \sqrt{T})$. Moreover, we show how the $\mathcal{NP}$-hard projection can be effectively executed on an adiabatic quantum annealer, harnessing recent advancements in quantum computation. We also introduce a projected version of this update rule and prove that if a fixed point exists in the binary variable space, the modified updates will converge to it. Last but not least, our algorithm is implemented layer-wise, making it suitable to train larger networks on resource-limited quantum hardware. Through extensive evaluations, we show that QP-SBGD outperforms or is on par with competitive and well-established baselines such as BinaryConnect, signSGD and ProxQuant when optimising the Rosenbrock function, training BNNs as well as binary graph neural networks.

Mixup is a popular data augmentation technique for training deep neural networks where additional samples are generated by linearly interpolating pairs of inputs and their labels. This technique is known to improve the generalization performance in many learning paradigms and applications. In this work, we first analyze Mixup and show that it implicitly regularizes infinitely many directional derivatives of all orders. Based on this new insight, we propose an improved version of Mixup, theoretically justified to deliver better generalization performance than the vanilla Mixup. To demonstrate the effectiveness of the proposed method, we conduct experiments across various domains such as images, tabular data, speech, and graphs. Our results show that the proposed method improves Mixup across multiple datasets using a variety of architectures, for instance, exhibiting an improvement over Mixup by 0.8% in ImageNet top-1 accuracy.

In many complex sequential decision-making tasks, online planning is crucial for high performance. For efficient online planning, Monte Carlo Tree Search (MCTS) employs a principled mechanism for trading off exploration for exploitation. MCTS outperforms comparison methods in many discrete decision-making domains such as Go, Chess, and Shogi. Following, extensions of MCTS to continuous domains have been proposed. However, the inherent high branching factor and the resulting explosion of search tree size are limiting existing methods. To address this problem, we propose Continuous Monte Carlo Graph Search (CMCGS), a novel extension of MCTS to online planning in environments with continuous state and action spaces. CMCGS takes advantage of the insight that, during planning, sharing the same action policy between several states can yield high performance. To implement this idea, at each time step, CMCGS clusters similar states into a limited number of stochastic action bandit nodes, which produce a layered directed graph instead of an MCTS search tree. Experimental evaluation shows that CMCGS outperforms comparable planning methods in several complex continuous DeepMind Control Suite benchmarks and a 2D navigation task with limited sample budgets. Furthermore, CMCGS can be parallelized to scale up and it outperforms the Cross-Entropy Method (CEM) in continuous control with learned dynamics models.

Reinforcement learning is able to solve complex sequential decision-making tasks but is currently limited by sample efficiency and required computation. To improve sample efficiency, recent work focuses on model-based RL which interleaves model learning with planning. Recent methods further utilize policy learning, value estimation, and, self-supervised learning as auxiliary objectives. In this paper we show that, surprisingly, a simple representation learning approach relying only on a latent dynamics model trained by latent temporal consistency is sufficient for high-performance RL. This applies when using pure planning with a dynamics model conditioned on the representation, but, also when utilizing the representation as policy and value function features in model-free RL. In experiments, our approach learns an accurate dynamics model to solve challenging high-dimensional locomotion tasks with online planners while being 4.1 times faster to train compared to ensemble-based methods. With model-free RL without planning, especially on high-dimensional tasks, such as the DeepMind Control Suite Humanoid and Dog tasks, our approach outperforms model-free methods by a large margin and matches model-based methods' sample efficiency while training 2.4 times faster.

High-definition (HD) semantic map generation of the environment is an essential component of autonomous driving. Existing methods have achieved good performance in this task by fusing different sensor modalities, such as LiDAR and camera. However, current works are based on raw data or network feature-level fusion and only consider short-range HD map generation, limiting their deployment to realistic autonomous driving applications. In this paper, we focus on the task of building the HD maps in both short ranges, i.e., within 30 m, and also predicting long-range HD maps up to 90 m, which is required by downstream path planning and control tasks to improve the smoothness and safety of autonomous driving. To this end, we propose a novel network named SuperFusion, exploiting the fusion of LiDAR and camera data at multiple levels. We use LiDAR depth to improve image depth estimation and use image features to guide long-range LiDAR feature prediction. We benchmark our SuperFusion on the nuScenes dataset and a self-recorded dataset and show that it outperforms the state-of-the-art baseline methods with large margins on all intervals. Additionally, we apply the generated HD map to a downstream path planning task, demonstrating that the long-range HD maps predicted by our method can lead to better path planning for autonomous vehicles. Our code and self-recorded dataset will be available at https://github.com/haomo-ai/SuperFusion.