强化学习

Synthesizing natural human motions that enable a 3D human avatar to walk and reach for arbitrary goals in 3D space remains an unsolved problem with many applications. Existing methods (data-driven or using reinforcement learning) are limited in terms of generalization and motion naturalness. A primary obstacle is the scarcity of training data that combines locomotion with goal reaching. To address this we introduce WANDR a data-driven model that takes an avatar's initial pose and a goal's 3D position and generates natural human motions that place the end effector (wrist) on the goal location. To solve this we introduce novel intention features that drive rich goal-oriented movement. Intention guides the agent to the goal and interactively adapts the generation to novel situations without needing to define sub-goals or the entire motion path. Crucially intention allows training on datasets that have goal-oriented motions as well as those that do not. WANDR is a conditional Variational Auto-Encoder (c-VAE) which we train using the AMASS and CIRCLE datasets. We evaluate our method extensively and demonstrate its ability to generate natural and long-term motions that reach 3D goals and generalize to unseen goal locations. Our models and code are available for research purposes at wandr.is.tue.mpg.de

Using reinforcement learning with human feedback (RLHF) has shown significant promise in fine-tuning diffusion models. Previous methods start by training a reward model that aligns with human preferences then leverage RL techniques to fine-tune the underlying models. However crafting an efficient reward model demands extensive datasets optimal architecture and manual hyperparameter tuning making the process both time and cost-intensive. The direct preference optimization (DPO) method effective in fine-tuning large language models eliminates the necessity for a reward model. However the extensive GPU memory requirement of the diffusion model's denoising process hinders the direct application of the DPO method. To address this issue we introduce the Direct Preference for Denoising Diffusion Policy Optimization (D3PO) method to directly fine-tune diffusion models. The theoretical analysis demonstrates that although D3PO omits training a reward model it effectively functions as the optimal reward model trained using human feedback data to guide the learning process. This approach requires no training of a reward model proving to be more direct cost-effective and minimizing computational overhead. In experiments our method uses the relative scale of objectives as a proxy for human preference delivering comparable results to methods using ground-truth rewards. Moreover D3PO demonstrates the ability to reduce image distortion rates and generate safer images overcoming challenges lacking robust reward models. Our code is publicly available at https://github.com/yk7333/D3PO.

Diffusion models have demonstrated unprecedented capabilities in image generation. Yet they incorporate and amplify the data bias (e.g. gender age) from the original training set limiting the diversity of generated images. In this paper we propose a diversity-oriented fine-tuning method using reinforcement learning (RL) for diffusion models under the guidance of an image-set-based reward function. Specifically the proposed reward function denoted as Diversity Reward utilizes a set of generated images to evaluate the coverage of the current generative distribution w.r.t. the reference distribution represented by a set of unbiased images. Built on top of the probabilistic method of distribution discrepancy estimation Diversity Reward can measure the relative distribution gap with a small set of images efficiently. We further formulate the diffusion process as a multi-step decision-making problem (MDP) and apply policy gradient methods to fine-tune diffusion models by maximizing the Diversity Reward. The proposed rewards are validated on a post-sampling selection task where a subset of the most diverse images are selected based on Diversity Reward values. We also show the effectiveness of our RL fine-tuning framework on enhancing the diversity of image generation with different types of diffusion models including class-conditional models and text-conditional models e.g. StableDiffusion.

Spike camera is a neuromorphic vision sensor that can capture highly dynamic scenes by generating a continuous stream of binary spikes to represent the arrival of photons at very high temporal resolution. Equipped with Bayer color filter array (CFA) color spike camera (CSC) has been invented to capture color information. Although spike camera has already demonstrated great potential for high-speed imaging its spatial resolution is limited compared with conventional digital cameras. This paper proposes a Color Spike Camera Super-Resolution (CSCSR) network to super-resolve higher-resolution color images from spike camera streams with Bayer CFA. To be specific we first propose a representation for Bayer-pattern spike streams exploring local temporal information with global perception to represent the binary data. Then we exploit the CFA layout and sub-pixel level motion to collect temporal pixels for the spatial super-resolution of each color channel. In particular a residual-based module for feature refinement is developed to reduce the impact of motion estimation errors. Considering color correlation we jointly utilize the multi-stage temporal-pixel features of color channels to reconstruct the high-resolution color image. Experimental results demonstrate that the proposed scheme can reconstruct satisfactory color images with both high temporal and spatial resolution from low-resolution Bayer-pattern spike streams. The source codes are available at https://github.com/csycdong/CSCSR.

We propose a novel self-supervised embedding to learn how actions sound from narrated in-the-wild egocentric videos. Whereas existing methods rely on curated data with known audio-visual correspondence our multimodal contrastive-consensus coding (MC3) embedding reinforces the associations between audio language and vision when all modality pairs agree while diminishing those associations when any one pair does not. We show our approach can successfully discover how the long tail of human actions sound from egocentric video outperforming an array of recent multimodal embedding techniques on two datasets (Ego4D and EPIC-Sounds) and multiple cross-modal tasks.

Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data avoiding privacy security and proprietary risks in real applications. In this line of research existing methods typically follow an inversion-and-distillation paradigm in which a generative adversarial network on-the-fly trained with the guidance of the pre-trained teacher network is used to synthesize a large-scale sample set for knowledge distillation. In this paper we reexamine this common data-free knowledge distillation paradigm showing that there is considerable room to improve the overall training efficiency through a lens of "small-scale inverted data for knowledge distillation". In light of three empirical observations indicating the importance of how to balance class distributions in terms of synthetic sample diversity and difficulty during both data inversion and distillation processes we propose Small Scale Data-free Knowledge Distillation (SSD-KD). In formulation SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples facilitated by a dynamic replay buffer and a reinforcement learning strategy. As a result SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples (e.g. 10x less than the original training data scale) making the overall training efficiency one or two orders of magnitude faster than many mainstream methods while retaining superior or competitive model performance as demonstrated on popular image classification and semantic segmentation benchmarks. The code is available at https://github.com/OSVAI/SSD-KD.

Many instruments performing optical and non-optical imaging and sensing such as Optical Coherence Tomography (OCT) Magnetic Resonance Imaging or Fourier-transform spectrometry produce digital signals containing modulations sine-like components which only after Fourier transformation give information about the structure or characteristics of the investigated object. Due to the fundamental physics-related limitations of such methods the distribution of these signal components is often nonlinear and when not properly compensated leads to the resolution precision or quality drop in the final image. Here we propose an innovative approach that has the potential to allow cleaning of the signal from the nonlinearities but most of all it now allows to switch the given order off leaving all others intact. The latter provides a tool for more in-depth analysis of the nonlinearity-inducing properties of the investigated object which can lead to applications in early disease detection or more sensitive sensing of chemical compounds. We consider OCT signals and nonlinearities up to the third order. In our approach we propose two neural networks: one to remove solely the second-order nonlinearity and the other for removing solely the third-order nonlinearity. The input of the networks is a novel two-dimensional data structure with all the information needed for the network to infer a nonlinearity-free signal. We describe the developed networks and present the results for second-order and third-order nonlinearity removal in OCT data representing the images of various objects: a mirror glass and fruits.

Reinforcement learning (RL) with dense rewards and imitation learning (IL) with human-generated trajectories are the most widely used approaches for training modern embodied agents. RL requires extensive reward shaping and auxiliary losses and is often too slow and ineffective for long-horizon tasks. While IL with human supervision is effective collecting human trajectories at scale is extremely expensive. In this work we show that imitating shortest-path planners in simulation produces agents that given a language instruction can proficiently navigate explore and manipulate objects in both simulation and in the real world using only RGB sensors (no depth map or GPS coordinates). This surprising result is enabled by our end-to-end transformer-based SPOC architecture powerful visual encoders paired with extensive image augmentation and the dramatic scale and diversity of our training data: millions of frames of shortest-path-expert trajectories collected inside approximately 200000 procedurally generated houses containing 40000 unique 3D assets. Our models data training code and newly proposed 10-task benchmarking suite CHORES are available at https://spoc-robot.github.io/.

The problem of self-calibration of two cameras from a given fundamental matrix is one of the basic problems in geometric computer vision. Under the assumption of known principal points and square pixels the Bougnoux formula offers a means to compute the two unknown focal lengths. However in many practical situations the formula yields inaccurate results due to commonly occurring singularities. Moreover the estimates are sensitive to noise in the computed fundamental matrix and to the assumed positions of the principal points. In this paper we therefore propose an efficient and robust iterative method to estimate the focal lengths along with the principal points of the cameras given a fundamental matrix and priors for the estimated camera intrinsics. In addition we study a computationally efficient check of models generated within RANSAC that improves the accuracy of the estimated models while reducing the total computational time. Extensive experiments on real and synthetic data show that our iterative method brings significant improvements in terms of the accuracy of the estimated focal lengths over the Bougnoux formula and other state-of-the-art methods even when relying on inaccurate priors. The code for the methods and experiments is available at https://github.com/kocurvik/robust_self_calibration

Recent Text-to-Image (T2I) generation models such as Stable Diffusion and Imagen have made significant progress in generating high-resolution images based on text descriptions. However many generated images still suffer from issues such as artifacts/implausibility misalignment with text descriptions and low aesthetic quality. Inspired by the success of Reinforcement Learning with Human Feedback (RLHF) for large language models prior works collected human-provided scores as feedback on generated images and trained a reward model to improve the T2I generation. In this paper we enrich the feedback signal by (i) marking image regions that are implausible or misaligned with the text and (ii) annotating which words in the text prompt are misrepresented or missing on the image. We collect such rich human feedback on 18K generated images (RichHF-18K) and train a multimodal transformer to predict the rich feedback automatically. We show that the predicted rich human feedback can be leveraged to improve image generation for example by selecting high-quality training data to finetune and improve the generative models or by creating masks with predicted heatmaps to inpaint the problematic regions. Notably the improvements generalize to models (Muse) beyond those used to generate the images on which human feedback data were collected (Stable Diffusion variants). The RichHF-18K data set will be released in our GitHub repository: https://github.com/google-research/google-research/tree/master/richhf_18k.

Recently lightweight Vision Transformers (ViTs) demonstrate superior performance and lower latency compared with lightweight Convolutional Neural Networks (CNNs) on resource-constrained mobile devices. Researchers have discovered many structural connections between lightweight ViTs and lightweight CNNs. However the notable architectural disparities in the block structure macro and micro designs between them have not been adequately examined. In this study we revisit the efficient design of lightweight CNNs from ViT perspective and emphasize their promising prospect for mobile devices. Specifically we incrementally enhance the mobile-friendliness of a standard lightweight CNN i.e. MobileNetV3 by integrating the efficient architectural designs of lightweight ViTs. This ends up with a new family of pure lightweight CNNs namely RepViT. Extensive experiments show that RepViT outperforms existing state-of-the-art lightweight ViTs and exhibits favorable latency in various vision tasks. Notably on ImageNet RepViT achieves over 80% top-1 accuracy with 1.0 ms latency on an iPhone 12 which is the first time for a lightweight model to the best of our knowledge. Besides when RepViT meets SAM our RepViT-SAM can achieve nearly 10xfaster inference than the advanced MobileSAM. Codes and models are available at https://github.com/THU-MIG/RepViT.

In order for reinforcement learning (RL) agents to be deployed in real-world environments they must be able to generalize to unseen environments. However RL struggles with out-of-distribution generalization often due to over-fitting the particulars of the training environment. Although regularization techniques from supervised learning can be applied to avoid over-fitting the differences between supervised learning and RL limit their application. To address this we propose the Signal-to-Noise Ratio regulated Parameter Uncertainty Network (SNR PUN) for RL. We introduce SNR as a new measure of regularizing the parameter uncertainty of a network and provide a formal analysis explaining why SNR regularization works well for RL. We demonstrate the effectiveness of our proposed method to generalize in several simulated environments; and in a physical system showing the possibility of using SNR PUN for applying RL to real-world applications.

While recent model-free Reinforcement Learning (RL) methods have demonstrated human-level effectiveness in gaming environments their success in everyday tasks like visual navigation has been limited particularly under significant appearance variations. This limitation arises from (i) poor sample efficiency and (ii) over-fitting to training scenarios. To address these challenges we present a world model that learns invariant features using (i) contrastive unsupervised learning and (ii) an intervention-invariant regularizer. Learning an explicit representation of the world dynamics i.e. a world model improves sample efficiency while contrastive learning implicitly enforces learning of invariant features which improves generalization. However the naive integration of contrastive loss to world models is not good enough as world-model-based RL methods independently optimize representation learning and agent policy. To overcome this issue we propose an intervention-invariant regularizer in the form of an auxiliary task such as depth prediction image denoising image segmentation etc. that explicitly enforces invariance to style interventions. Our method outperforms current state-of-the-art model-based and model-free RL methods and significantly improves on out-of-distribution point navigation tasks evaluated on the iGibson benchmark. With only visual observations we further demonstrate that our approach outperforms recent language-guided foundation models for point navigation which is essential for deployment on robots with limited computation capabilities. Finally we demonstrate that our proposed model excels at the sim-to-real transfer of its perception module on the Gibson benchmark.

We leverage Large Language Models (LLM) for zeroshot Semantic Audio Visual Navigation (SAVN). Existing methods utilize extensive training demonstrations for reinforcement learning yet achieve relatively low success rates and lack generalizability. The intermittent nature of auditory signals further poses additional obstacles to inferring the goal information. To address this challenge we present the Reflective and Imaginative Language Agent (RILA). By employing multi-modal models to process sensory data we instruct an LLM-based planner to actively explore the environment. During the exploration our agent adaptively evaluates and dismisses inaccurate perceptual descriptions. Additionally we introduce an auxiliary LLMbased assistant to enhance global environmental comprehension by mapping room layouts and providing strategic insights. Through comprehensive experiments and analysis we show that our method outperforms relevant baselines without training demonstrations from the environment and complementary semantic information.

Customizing robotic behaviors to be aligned with diverse human preferences is an underexplored challenge in the field of embodied AI. In this paper we present Promptable Behaviors a novel framework that facilitates efficient personalization of robotic agents to diverse human preferences in complex environments. We use multi-objective reinforcement learning to train a single policy adaptable to a broad spectrum of preferences. We introduce three distinct methods to infer human preferences by leveraging different types of interactions: (1) human demonstrations (2) preference feedback on trajectory comparisons and (3) language instructions. We evaluate the proposed method in personalized object-goal navigation and flee navigation tasks in ProcTHOR and RoboTHOR demonstrating the ability to prompt agent behaviors to satisfy human preferences in various scenarios.

Infrared and visible image fusion aims to generate a fused image by integrating and distinguishing complementary information from multiple sources. While the cross-attention mechanism with global spatial interactions appears promising it only capture second-order spatial interactions neglecting higher-order interactions in both spatial and channel dimensions. This limitation hampers the exploitation of synergies between multi-modalities. To bridge this gap we introduce a Synergistic High-order Interaction Paradigm (SHIP) designed to systematically investigate spatial fine-grained and global statistics collaborations between infrared and visible images across two fundamental dimensions: 1) Spatial dimension: we construct spatial fine-grained interactions through element-wise multiplication mathematically equivalent to global interactions and then foster high-order formats by iteratively aggregating and evolving complementary information enhancing both efficiency and flexibility. 2) Channel dimension: expanding on channel interactions with first-order statistics (mean) we devise high-order channel interactions to facilitate the discernment of inter-dependencies between source images based on global statistics. Harnessing high-order interactions significantly enhances our model's ability to exploit multi-modal synergies leading in superior performance over state-of-the-art alternatives as shown through comprehensive experiments across various benchmarks.

We introduce multimodal story summarization by leveraging TV episode recaps - short video sequences interweaving key story moments from previous episodes to bring viewers up to speed. We propose PlotSnap a dataset featuring two crime thriller TV shows with rich recaps and long episodes of 40 minutes. Story summarization labels are unlocked by matching recap shots to corresponding sub-stories in the episode. We propose a hierarchical model TaleSumm that processes entire episodes by creating compact shot and dialog representations and predicts importance scores for each video shot and dialog utterance by enabling interactions between local story groups. Unlike traditional summarization our method extracts multiple plot points from long videos. We present a thorough evaluation on story summarization including promising cross-series generalization. TaleSumm also shows good results on classic video summarization benchmarks.

Computer-Aided Design (CAD) model reconstruction from point clouds is an important problem at the intersection of computer vision graphics and machine learning; it saves the designer significant time when iterating on in-the-wild objects. Recent advancements in this direction achieve relatively reliable semantic segmentation but still struggle to produce an adequate topology of the CAD model. In this work we analyze the current state of the art for that ill-posed task and identify shortcomings of existing methods. We propose a hybrid analytic-neural reconstruction scheme that bridges the gap between segmented point clouds and structured CAD models and can be readily combined with different segmentation backbones. Moreover to power the surface fitting stage we propose a novel implicit neural representation of freeform surfaces driving up the performance of our overall CAD reconstruction scheme. We extensively evaluate our method on the popular ABC benchmark of CAD models and set a new state-of-the-art for that dataset. Code is available at https://github.com/YujiaLiu76/point2cad.

Reward finetuning has emerged as a promising approach to aligning foundation models with downstream objectives. Remarkable success has been achieved in the language domain by using reinforcement learning (RL) to maximize rewards that reflect human preference. However in the vision domain existing RL-based reward finetuning methods are limited by their instability in large-scale training rendering them incapable of generalizing to complex unseen prompts. In this paper we propose Proximal Reward Difference Prediction (PRDP) enabling stable black-box reward finetuning for diffusion models for the first time on large-scale prompt datasets with over 100K prompts. Our key innovation is the Reward Difference Prediction (RDP) objective that has the same optimal solution as the RL objective while enjoying better training stability. Specifically the RDP objective is a supervised regression objective that tasks the diffusion model with predicting the reward difference of generated image pairs from their denoising trajectories. We theoretically prove that the diffusion model that obtains perfect reward difference prediction is exactly the maximizer of the RL objective. We further develop an online algorithm with proximal updates to stably optimize the RDP objective. In experiments we demonstrate that PRDP can match the reward maximization ability of well-established RL-based methods in small-scale training. Furthermore through large-scale training on text prompts from the Human Preference Dataset v2 and the Pick-a-Pic v1 dataset PRDP achieves superior generation quality on a diverse set of complex unseen prompts whereas RL-based methods completely fail.

Multi-constraint offline reinforcement learning (RL) promises to learn policies that satisfy both cumulative and state-wise costs from offline datasets. This arrangement provides an effective approach for the widespread application of RL in high-risk scenarios where both cumulative and state-wise costs need to be considered simultaneously. However previously constrained offline RL algorithms are primarily designed to handle single-constraint problems related to cumulative cost which faces challenges when addressing multi-constraint tasks that involve both cumulative and state-wise costs. In this work we propose a novel Primal policy Optimization with Conservative Estimation algorithm (POCE) to address the problem of multi-constraint offline RL. Concretely we reframe the objective of multi-constraint offline RL by introducing the concept of Maximum Markov Decision Processes (MMDP). Subsequently we present a primal policy optimization algorithm to confront the multi-constraint problems which improves the stability and convergence speed of model training. Furthermore we propose a conditional Bellman operator to estimate cumulative and state-wise Q-values reducing the extrapolation error caused by out-of-distribution (OOD) actions. Finally extensive experiments demonstrate that the POCE algorithm achieves competitive performance across multiple experimental tasks particularly outperforming baseline algorithms in terms of safety. Our code is available at \href https://github.com/guanjiayi/poce github.POCE .