A Conversational Recommender System (CRS) offers increased transparency and control to users by enabling them to engage with the system through a real-time multi-turn dialogue. Recently, Large Language Models (LLMs) have exhibited an unprecedented ability to converse naturally and incorporate world knowledge and common-sense reasoning into language understanding, unlocking the potential of this paradigm. However, effectively leveraging LLMs within a CRS introduces new technical challenges, including properly understanding and controlling a complex conversation and retrieving from external sources of information. These issues are exacerbated by a large, evolving item corpus and a lack of conversational data for training. In this paper, we provide a roadmap for building an end-to-end large-scale CRS using LLMs. In particular, we propose new implementations for user preference understanding, flexible dialogue management and explainable recommendations as part of an integrated architecture powered by LLMs. For improved personalization, we describe how an LLM can consume interpretable natural language user profiles and use them to modulate session-level context. To overcome conversational data limitations in the absence of an existing production CRS, we propose techniques for building a controllable LLM-based user simulator to generate synthetic conversations. As a proof of concept we introduce RecLLM, a large-scale CRS for YouTube videos built on LaMDA, and demonstrate its fluency and diverse functionality through some illustrative example conversations.
Pose registration is critical in vision and robotics. This paper focuses on the challenging task of initialization-free pose registration up to 7DoF for homogeneous and heterogeneous measurements. While recent learning-based methods show promise using differentiable solvers, they either rely on heuristically defined correspondences or are prone to local minima. We present a differentiable phase correlation (DPC) solver that is globally convergent and correspondence-free. When combined with simple feature extraction networks, our general framework DPCN++ allows for versatile pose registration with arbitrary initialization. Specifically, the feature extraction networks first learn dense feature grids from a pair of homogeneous/heterogeneous measurements. These feature grids are then transformed into a translation and scale invariant spectrum representation based on Fourier transform and spherical radial aggregation, decoupling translation and scale from rotation. Next, the rotation, scale, and translation are independently and efficiently estimated in the spectrum step-by-step using the DPC solver. The entire pipeline is differentiable and trained end-to-end. We evaluate DCPN++ on a wide range of registration tasks taking different input modalities, including 2D bird's-eye view images, 3D object and scene measurements, and medical images. Experimental results demonstrate that DCPN++ outperforms both classical and learning-based baselines, especially on partially observed and heterogeneous measurements.
Depth completion is a fundamental task in computer vision and robotics research. Many previous works complete the dense depth map with neural networks directly but most of them are non-interpretable and can not generalize to different situations well. In this paper, we propose an effective image representation method for depth completion tasks. The input of our system is a monocular camera frame and the synchronous sparse depth map. The output of our system is a dense per-pixel depth map of the frame. First we use a neural network to transform each pixel into a feature vector, which we call base functions. Then we pick out the known pixels' base functions and their depth values. We use a linear least square algorithm to fit the base functions and the depth values. Then we get the weights estimated from the least square algorithm. Finally, we apply the weights to the whole image and predict the final depth map. Our method is interpretable so it can generalize well. Experiments show that our results beat the state-of-the-art on NYU-Depth-V2 dataset both in accuracy and runtime. Moreover, experiments show that our method can generalize well on different numbers of sparse points and different datasets.
Monocular visual-inertial odometry (VIO) is a critical problem in robotics and autonomous driving. Traditional methods solve this problem based on filtering or optimization. While being fully interpretable, they rely on manual interference and empirical parameter tuning. On the other hand, learning-based approaches allow for end-to-end training but require a large number of training data to learn millions of parameters. However, the non-interpretable and heavy models hinder the generalization ability. In this paper, we propose a fully differentiable, interpretable, and lightweight monocular VIO model that contains only 4 trainable parameters. Specifically, we first adopt Unscented Kalman Filter as a differentiable layer to predict the pitch and roll, where the covariance matrices of noise are learned to filter out the noise of the IMU raw data. Second, the refined pitch and roll are adopted to retrieve a gravity-aligned BEV image of each frame using differentiable camera projection. Finally, a differentiable pose estimator is utilized to estimate the remaining 4 DoF poses between the BEV frames. Our method allows for learning the covariance matrices end-to-end supervised by the pose estimation loss, demonstrating superior performance to empirical baselines. Experimental results on synthetic and real-world datasets demonstrate that our simple approach is competitive with state-of-the-art methods and generalizes well on unseen scenes.
One of the challenges in vision-based driving trajectory generation is dealing with out-of-distribution scenarios. In this paper, we propose a domain generalization method for vision-based driving trajectory generation for autonomous vehicles in urban environments, which can be seen as a solution to extend the Invariant Risk Minimization (IRM) method in complex problems. We leverage an adversarial learning approach to train a trajectory generator as the decoder. Based on the pre-trained decoder, we infer the latent variables corresponding to the trajectories, and pre-train the encoder by regressing the inferred latent variable. Finally, we fix the decoder but fine-tune the encoder with the final trajectory loss. We compare our proposed method with the state-of-the-art trajectory generation method and some recent domain generalization methods on both datasets and simulation, demonstrating that our method has better generalization ability.
Motion retargeting from human to robot remains a very challenging task due to variations in the structure of humans and robots. Most traditional optimization-based algorithms solve this problem by minimizing an objective function, which is usually time-consuming and heavily dependent on good initialization. In contrast, methods with feedforward neural networks can learn prior knowledge from training data and quickly infer the results, but these methods also suffer from the generalization problem on unseen actions, leading to some infeasible results. In this paper, we propose a novel neural optimization approach taking advantages of both kinds of methods. A graph-based neural network is utilized to establish a mapping between the latent space and the robot motion space. Afterward, the retargeting results can be obtained by searching for the optimal vector in this latent space. In addition, a deep encoder also provides a promising initialization for better and faster convergence. We perform experiments on retargeting Chinese sign language to three different kinds of robots in the simulation environment, including ABB's YuMi dual-arm collaborative robot, NAO and Pepper. A real-world experiment is also conducted on the Yumi robot. Experimental results show that our method can retarget motion from human to robot with both efficiency and accuracy.
Recent studies on deep-learning-based small defection segmentation approaches are trained in specific settings and tend to be limited by fixed context. Throughout the training, the network inevitably learns the representation of the background of the training data before figuring out the defection. They underperform in the inference stage once the context changed and can only be solved by training in every new setting. This eventually leads to the limitation in practical robotic applications where contexts keep varying. To cope with this, instead of training a network context by context and hoping it to generalize, why not stop misleading it with any limited context and start training it with pure simulation? In this paper, we propose the network SSDS that learns a way of distinguishing small defections between two images regardless of the context, so that the network can be trained once for all. A small defection detection layer utilizing the pose sensitivity of phase correlation between images is introduced and is followed by an outlier masking layer. The network is trained on randomly generated simulated data with simple shapes and is generalized across the real world. Finally, SSDS is validated on real-world collected data and demonstrates the ability that even when trained in cheap simulation, SSDS can still find small defections in the real world showing the effectiveness and its potential for practical applications.
Utilizing the trained model under different conditions without data annotation is attractive for robot applications. Towards this goal, one class of methods is to translate the image style from the training environment to the current one. Conventional studies on image style translation mainly focus on two settings: paired data on images from two domains with exactly aligned content, and unpaired data, with independent content. In this paper, we would like to propose a new setting, where the content in the two images is aligned with error in poses. We consider that this setting is more practical since robots with various sensors are able to align the data up to some error level, even with different styles. To solve this problem, we propose PRoGAN to learn a style translator by intentionally transforming the original domain images with a noisy pose, then matching the distribution of translated transformed images and the distribution of the target domain images. The adversarial training enforces the network to learn the style translation, avoiding being entangled with other variations. In addition, we propose two pose estimation based self-supervised tasks to further improve the performance. Finally, PRoGAN is validated on both simulated and real-world collected data to show the effectiveness. Results on down-stream tasks, classification, road segmentation, object detection, and feature matching show its potential for real applications. https://github.com/wrld/PRoGAN .
Global localization is essential for robot navigation, of which the first step is to retrieve a query from the map database. This problem is called place recognition. In recent years, LiDAR scan based place recognition has drawn attention as it is robust against the environmental change. In this paper, we propose a LiDAR-based place recognition method, named Differentiable Scan Context with Orientation (DiSCO), which simultaneously finds the scan at a similar place and estimates their relative orientation. The orientation can further be used as the initial value for the down-stream local optimal metric pose estimation, improving the pose estimation especially when a large orientation between the current scan and retrieved scan exists. Our key idea is to transform the feature learning into the frequency domain. We utilize the magnitude of the spectrum as the place signature, which is theoretically rotation-invariant. In addition, based on the differentiable phase correlation, we can efficiently estimate the global optimal relative orientation using the spectrum. With such structural constraints, the network can be learned in an end-to-end manner, and the backbone is fully shared by the two tasks, achieving interpretability and light weight. Finally, DiSCO is validated on the NCLT and Oxford datasets with long-term outdoor conditions, showing better performance than the compared methods.