Yang Gao

Renhao Wang, Jiayuan Mao, Joy Hsu et al. · stanford

Robots operating in the real world require both rich manipulation skills as well as the ability to semantically reason about when to apply those skills. Towards this goal, recent works have integrated semantic representations from large-scale pretrained vision-language (VL) models into manipulation models, imparting them with more general reasoning capabilities. However, we show that the conventional pretraining-finetuning pipeline for integrating such representations entangles the learning of domain-specific action information and domain-general visual information, leading to less data-efficient training and poor generalization to unseen objects and tasks. To this end, we propose ProgramPort, a modular approach to better leverage pretrained VL models by exploiting the syntactic and semantic structures of language instructions. Our framework uses a semantic parser to recover an executable program, composed of functional modules grounded on vision and action across different modalities. Each functional module is realized as a combination of deterministic computation and learnable neural networks. Program execution produces parameters to general manipulation primitives for a robotic end-effector. The entire modular network can be trained with end-to-end imitation learning objectives. Experiments show that our model successfully disentangles action and perception, translating to improved zero-shot and compositional generalization in a variety of manipulation behaviors. Project webpage at: \url{https://progport.github.io}.

Heng Cai, Lei Qi, Qian Yu et al.

Medical image segmentation typically necessitates a large and precisely annotated dataset. However, obtaining pixel-wise annotation is a labor-intensive task that requires significant effort from domain experts, making it challenging to obtain in practical clinical scenarios. In such situations, reducing the amount of annotation required is a more practical approach. One feasible direction is sparse annotation, which involves annotating only a few slices, and has several advantages over traditional weak annotation methods such as bounding boxes and scribbles, as it preserves exact boundaries. However, learning from sparse annotation is challenging due to the scarcity of supervision signals. To address this issue, we propose a framework that can robustly learn from sparse annotation using the cross-teaching of both 3D and 2D networks. Considering the characteristic of these networks, we develop two pseudo label selection strategies, which are hard-soft confidence threshold and consistent label fusion. Our experimental results on the MMWHS dataset demonstrate that our method outperforms the state-of-the-art (SOTA) semi-supervised segmentation methods. Moreover, our approach achieves results that are comparable to the fully-supervised upper bound result.

Yingdong Hu, Fanqi Lin, Tong Zhang et al.

In this study, we are interested in imbuing robots with the capability of physically-grounded task planning. Recent advancements have shown that large language models (LLMs) possess extensive knowledge useful in robotic tasks, especially in reasoning and planning. However, LLMs are constrained by their lack of world grounding and dependence on external affordance models to perceive environmental information, which cannot jointly reason with LLMs. We argue that a task planner should be an inherently grounded, unified multimodal system. To this end, we introduce Robotic Vision-Language Planning (ViLa), a novel approach for long-horizon robotic planning that leverages vision-language models (VLMs) to generate a sequence of actionable steps. ViLa directly integrates perceptual data into its reasoning and planning process, enabling a profound understanding of commonsense knowledge in the visual world, including spatial layouts and object attributes. It also supports flexible multimodal goal specification and naturally incorporates visual feedback. Our extensive evaluation, conducted in both real-robot and simulated environments, demonstrates ViLa's superiority over existing LLM-based planners, highlighting its effectiveness in a wide array of open-world manipulation tasks.

Tong Zhang, Yingdong Hu, Hanchen Cui et al.

Robots rely heavily on sensors, especially RGB and depth cameras, to perceive and interact with the world. RGB cameras record 2D images with rich semantic information while missing precise spatial information. On the other side, depth cameras offer critical 3D geometry data but capture limited semantics. Therefore, integrating both modalities is crucial for learning representations for robotic perception and control. However, current research predominantly focuses on only one of these modalities, neglecting the benefits of incorporating both. To this end, we present $\textbf{Semantic-Geometric Representation} (\textbf{SGR})$, a universal perception module for robotics that leverages the rich semantic information of large-scale pre-trained 2D models and inherits the merits of 3D spatial reasoning. Our experiments demonstrate that SGR empowers the agent to successfully complete a diverse range of simulated and real-world robotic manipulation tasks, outperforming state-of-the-art methods significantly in both single-task and multi-task settings. Furthermore, SGR possesses the capability to generalize to novel semantic attributes, setting it apart from the other methods. Project website: https://semantic-geometric-representation.github.io.

Yu Bai, Xiyuan Zou, Heyan Huang et al.

Long sequence modeling has gained broad interest as large language models (LLMs) continue to advance. Recent research has identified that a large portion of hidden states within the key-value caches of Transformer models can be discarded (also termed evicted) without affecting the perplexity performance in generating long sequences. However, we show that these methods, despite preserving perplexity performance, often drop information that is important for solving downstream tasks, a problem which we call information neglect. To address this issue, we introduce Chunked Instruction-aware State Eviction (CItruS), a novel modeling technique that integrates the attention preferences useful for a downstream task into the eviction process of hidden states. In addition, we design a method for chunked sequence processing to further improve efficiency. Our training-free method exhibits superior performance on long sequence comprehension and retrieval tasks over several strong baselines under the same memory budget, while preserving language modeling perplexity. The code and data have been released at https://github.com/ybai-nlp/CItruS.

Shengjie Wang, Shaohuai Liu, Weirui Ye et al.

Sample efficiency remains a crucial challenge in applying Reinforcement Learning (RL) to real-world tasks. While recent algorithms have made significant strides in improving sample efficiency, none have achieved consistently superior performance across diverse domains. In this paper, we introduce EfficientZero V2, a general framework designed for sample-efficient RL algorithms. We have expanded the performance of EfficientZero to multiple domains, encompassing both continuous and discrete actions, as well as visual and low-dimensional inputs. With a series of improvements we propose, EfficientZero V2 outperforms the current state-of-the-art (SOTA) by a significant margin in diverse tasks under the limited data setting. EfficientZero V2 exhibits a notable advancement over the prevailing general algorithm, DreamerV3, achieving superior outcomes in 50 of 66 evaluated tasks across diverse benchmarks, such as Atari 100k, Proprio Control, and Vision Control.

Rujia Yang, Geng Chen, Chuan Wen et al.

Following its success in natural language processing and computer vision, foundation models that are pre-trained on large-scale multi-task datasets have also shown great potential in robotics. However, most existing robot foundation models rely solely on 2D image observations, ignoring 3D geometric information, which is essential for robots to perceive and reason about the 3D world. In this paper, we introduce FP3, a first large-scale 3D foundation policy model for robotic manipulation. FP3 builds on a scalable diffusion transformer architecture and is pre-trained on 60k trajectories with point cloud observations. With the model design and diverse pre-training data, FP3 can be efficiently fine-tuned for downstream tasks while exhibiting strong generalization capabilities. Experiments on real robots demonstrate that with only 80 demonstrations, FP3 is able to learn a new task with over 90% success rates in novel environments with unseen objects, significantly surpassing existing robot foundation models.

Hongye Cao, Fan Feng, Meng Fang et al.

In Model-Based Reinforcement Learning (MBRL), incorporating causal structures into dynamics models provides agents with a structured understanding of the environments, enabling efficient decision. Empowerment as an intrinsic motivation enhances the ability of agents to actively control their environments by maximizing the mutual information between future states and actions. We posit that empowerment coupled with causal understanding can improve controllability, while enhanced empowerment gain can further facilitate causal reasoning in MBRL. To improve learning efficiency and controllability, we propose a novel framework, Empowerment through Causal Learning (ECL), where an agent with the awareness of causal dynamics models achieves empowerment-driven exploration and optimizes its causal structure for task learning. Specifically, ECL operates by first training a causal dynamics model of the environment based on collected data. We then maximize empowerment under the causal structure for exploration, simultaneously using data gathered through exploration to update causal dynamics model to be more controllable than dense dynamics model without causal structure. In downstream task learning, an intrinsic curiosity reward is included to balance the causality, mitigating overfitting. Importantly, ECL is method-agnostic and is capable of integrating various causal discovery methods. We evaluate ECL combined with 3 causal discovery methods across 6 environments including pixel-based tasks, demonstrating its superior performance compared to other causal MBRL methods, in terms of causal discovery, sample efficiency, and asymptotic performance.

Shenyang Liu, Yang Gao, Shaoyan Zhai et al.

Prompt Recovery, reconstructing prompts from the outputs of large language models (LLMs), has grown in importance as LLMs become ubiquitous. Most users access LLMs through APIs without internal model weights, relying only on outputs and logits, which complicates recovery. This paper explores a unique prompt recovery task focused on reconstructing prompts for style transfer and rephrasing, rather than typical question-answering. We introduce a dataset created with LLM assistance, ensuring quality through multiple techniques, and test methods like zero-shot, few-shot, jailbreak, chain-of-thought, fine-tuning, and a novel canonical-prompt fallback for poor-performing cases. Our results show that one-shot and fine-tuning yield the best outcomes but highlight flaws in traditional sentence similarity metrics for evaluating prompt recovery. Contributions include (1) a benchmark dataset, (2) comprehensive experiments on prompt recovery strategies, and (3) identification of limitations in current evaluation metrics, all of which advance general prompt recovery research, where the structure of the input prompt is unrestricted.

Daoyuan Zhou, Xuchuang Wang, Lin Yang et al.

We study the stochastic Multiplayer Multi-Armed Bandit (MMAB) problem, where multiple players select arms to maximize their cumulative rewards. Collisions occur when two or more players select the same arm, resulting in no reward, and are observed by the players involved. We consider a distributed setting without central coordination, where each player can only observe their own actions and collision feedback. We propose a distributed algorithm with an adaptive, efficient communication protocol. The algorithm achieves near-optimal group and individual regret, with a communication cost of only $\mathcal{O}(\log\log T)$. Our experiments demonstrate significant performance improvements over existing baselines. Compared to state-of-the-art (SOTA) methods, our approach achieves a notable reduction in individual regret. Finally, we extend our approach to a periodic asynchronous setting, proving the lower bound for this problem and presenting an algorithm that achieves logarithmic regret.

Jialei Huang, Shuo Wang, Fanqi Lin et al.

Vision-Language-Action (VLA) models have shown remarkable achievements, driven by the rich implicit knowledge of their vision-language components. However, achieving generalist robotic agents demands precise grounding into physical interactions, especially in contact-rich scenarios where fine-grained force control is essential. We advance VLAs' implicit knowledge beyond identifying what to do, towards guiding how to physically interact with real world. This paper introduces Tactile-VLA, a novel framework that deeply fuses vision, language, action, and tactile sensing. This framework incorporates a hybrid position-force controller to translate the model's intentions into precise physical actions and a reasoning module that allows the robot to adapt its strategy based on tactile feedback. Experiments demonstrate Tactile-VLA's effectiveness and generalizability in three key aspects: (1) enabling tactile-aware instruction following, (2) utilizing tactile-relevant commonsense, and (3) facilitating adaptive tactile-involved reasoning. A key finding is that the VLM's prior knowledge already contains semantic understanding of physical interaction; by connecting it to the robot's tactile sensors with only a few demonstrations, we can activate this prior knowledge to achieve zero-shot generalization in contact-rich tasks.

Heyan Huang, Yinghao Li, Huashan Sun et al.

Recent studies have demonstrated that In-Context Learning (ICL), through the use of specific demonstrations, can align Large Language Models (LLMs) with human preferences known as In-Context Alignment (ICA), indicating that models can comprehend human instructions without requiring parameter adjustments. However, the exploration of the mechanism and applicability of ICA remains limited. In this paper, we begin by dividing the context text used in ICA into three categories: format, system prompt, and example. Through ablation experiments, we investigate the effectiveness of each part in enabling ICA to function effectively. We then examine how variants in these parts impact the model's alignment performance. Our findings indicate that the example part is crucial for enhancing the model's alignment capabilities, with changes in examples significantly affecting alignment performance. We also conduct a comprehensive evaluation of ICA's zero-shot capabilities in various alignment tasks. The results indicate that compared to parameter fine-tuning methods, ICA demonstrates superior performance in knowledge-based tasks and tool-use tasks. However, it still exhibits certain limitations in areas such as multi-turn dialogues and instruction following.

Chengbo Yuan, Chuan Wen, Tong Zhang et al.

We address the challenge of acquiring real-world manipulation skills with a scalable framework. We hold the belief that identifying an appropriate prediction target capable of leveraging large-scale datasets is crucial for achieving efficient and universal learning. Therefore, we propose to utilize 3D flow, which represents the future trajectories of 3D points on objects of interest, as an ideal prediction target. To exploit scalable data resources, we turn our attention to human videos. We develop, for the first time, a language-conditioned 3D flow prediction model directly from large-scale RGBD human video datasets. Our predicted flow offers actionable guidance, thus facilitating zero-shot skill transfer in real-world scenarios. We deploy our method with a policy based on closed-loop flow prediction. Remarkably, without any in-domain finetuning, our method achieves an impressive 81\% success rate in zero-shot human-to-robot skill transfer, covering 18 tasks in 6 scenes. Our framework features the following benefits: (1) scalability: leveraging cross-embodiment data resources; (2) wide application: multiple object categories, including rigid, articulated, and soft bodies; (3) stable skill transfer: providing actionable guidance with a small inference domain-gap. Code, data, and supplementary materials are available https://general-flow.github.io

Guang Yang, Xingguo Chen, Shangdong Yang et al.

With rising uncertainty in the real world, online Reinforcement Learning (RL) has been receiving increasing attention due to its fast learning capability and improving data efficiency. However, online RL often suffers from complex Value Function Approximation (VFA) and catastrophic interference, creating difficulty for the deep neural network to be applied to an online RL algorithm in a fully online setting. Therefore, a simpler and more adaptive approach is introduced to evaluate value function with the kernel-based model. Sparse representations are superior at handling interference, indicating that competitive sparse representations should be learnable, non-prior, non-truncated and explicit when compared with current sparse representation methods. Moreover, in learning sparse representations, attention mechanisms are utilized to represent the degree of sparsification, and a smooth attentive function is introduced into the kernel-based VFA. In this paper, we propose an Online Attentive Kernel-Based Temporal Difference (OAKTD) algorithm using two-timescale optimization and provide convergence analysis of our proposed algorithm. Experimental evaluations showed that OAKTD outperformed several Online Kernel-based Temporal Difference (OKTD) learning algorithms in addition to the Temporal Difference (TD) learning algorithm with Tile Coding on public Mountain Car, Acrobot, CartPole and Puddle World tasks.

Marco Visca, Sampo Kuutti, Roger Powell et al.

Terrain traversability analysis plays a major role in ensuring safe robotic navigation in unstructured environments. However, real-time constraints frequently limit the accuracy of online tests especially in scenarios where realistic robot-terrain interactions are complex to model. In this context, we propose a deep learning framework trained in an end-to-end fashion from elevation maps and trajectories to estimate the occurrence of failure events. The network is first trained and tested in simulation over synthetic maps generated by the OpenSimplex algorithm. The prediction performance of the Deep Learning framework is illustrated by being able to retain over 94% recall of the original simulator at 30% of the computational time. Finally, the network is transferred and tested on real elevation maps collected by the SEEKER consortium during the Martian rover test trial in the Atacama desert in Chile. We show that transferring and fine-tuning of an application-independent pre-trained model retains better performance than training uniquely on scarcely available real data.

Marco Visca, Arthur Bouton, Roger Powell et al.

Driving energy consumption plays a major role in the navigation of mobile robots in challenging environments, especially if they are left to operate unattended under limited on-board power. This paper reports on first results of an energy-aware path planner, which can provide estimates of the driving energy consumption and energy recovery of a robot traversing complex uneven terrains. Energy is estimated over trajectories making use of a self-supervised learning approach, in which the robot autonomously learns how to correlate perceived terrain point clouds to energy consumption and recovery. A novel feature of the method is the use of 1D convolutional neural network to analyse the terrain sequentially in the same temporal order as it would be experienced by the robot when moving. The performance of the proposed approach is assessed in simulation over several digital terrain models collected from real natural scenarios, and is compared with a heuristic inclination-based energy model. We show evidence of the benefit of our method to increase the overall prediction r2 score by 66.8% and to reduce the driving energy consumption over planned paths by 5.5%.

Zhicheng Zhou, Cheng Zhao, Daniel Adolfsson et al.

3D point cloud-based place recognition is highly demanded by autonomous driving in GPS-challenged environments and serves as an essential component (i.e. loop-closure detection) in lidar-based SLAM systems. This paper proposes a novel approach, named NDT-Transformer, for realtime and large-scale place recognition using 3D point clouds. Specifically, a 3D Normal Distribution Transform (NDT) representation is employed to condense the raw, dense 3D point cloud as probabilistic distributions (NDT cells) to provide the geometrical shape description. Then a novel NDT-Transformer network learns a global descriptor from a set of 3D NDT cell representations. Benefiting from the NDT representation and NDT-Transformer network, the learned global descriptors are enriched with both geometrical and contextual information. Finally, descriptor retrieval is achieved using a query-database for place recognition. Compared to the state-of-the-art methods, the proposed approach achieves an improvement of 7.52% on average top 1 recall and 2.73% on average top 1% recall on the Oxford Robotcar benchmark.

Pedro F. Proença, Yang Gao

This paper presents CAPE, a method to extract planes and cylinder segments from organized point clouds, which processes 640x480 depth images on a single CPU core at an average of 300 Hz, by operating on a grid of planar cells. While, compared to state-of-the-art plane extraction, the latency of CAPE is more consistent and 4-10 times faster, depending on the scene, we also demonstrate empirically that applying CAPE to visual odometry can improve trajectory estimation on scenes made of cylindrical surfaces (e.g. tunnels), whereas using a plane extraction approach that is not curve-aware deteriorates performance on these scenes. To use these geometric primitives in visual odometry, we propose extending a probabilistic RGB-D odometry framework based on points, lines and planes to cylinder primitives. Following this framework, CAPE runs on fused depth maps and the parameters of cylinders are modelled probabilistically to account for uncertainty and weight accordingly the pose optimization residuals.

Pedro F. Proença, Yang Gao

Active depth cameras suffer from several limitations, which cause incomplete and noisy depth maps, and may consequently affect the performance of RGB-D Odometry. To address this issue, this paper presents a visual odometry method based on point and line features that leverages both measurements from a depth sensor and depth estimates from camera motion. Depth estimates are generated continuously by a probabilistic depth estimation framework for both types of features to compensate for the lack of depth measurements and inaccurate feature depth associations. The framework models explicitly the uncertainty of triangulating depth from both point and line observations to validate and obtain precise estimates. Furthermore, depth measurements are exploited by propagating them through a depth map registration module and using a frame-to-frame motion estimation method that considers 3D-to-2D and 2D-to-3D reprojection errors, independently. Results on RGB-D sequences captured on large indoor and outdoor scenes, where depth sensor limitations are critical, show that the combination of depth measurements and estimates through our approach is able to overcome the absence and inaccuracy of depth measurements.

Pedro F. Proenca, Yang Gao

This work proposes a robust visual odometry method for structured environments that combines point features with line and plane segments, extracted through an RGB-D camera. Noisy depth maps are processed by a probabilistic depth fusion framework based on Mixtures of Gaussians to denoise and derive the depth uncertainty, which is then propagated throughout the visual odometry pipeline. Probabilistic 3D plane and line fitting solutions are used to model the uncertainties of the feature parameters and pose is estimated by combining the three types of primitives based on their uncertainties. Performance evaluation on RGB-D sequences collected in this work and two public RGB-D datasets: TUM and ICL-NUIM show the benefit of using the proposed depth fusion framework and combining the three feature-types, particularly in scenes with low-textured surfaces, dynamic objects and missing depth measurements.

Pedro F. Proença, Yang Gao

This work proposes a visual odometry method that combines points and plane primitives, extracted from a noisy depth camera. Depth measurement uncertainty is modelled and propagated through the extraction of geometric primitives to the frame-to-frame motion estimation, where pose is optimized by weighting the residuals of 3D point and planes matches, according to their uncertainties. Results on an RGB-D dataset show that the combination of points and planes, through the proposed method, is able to perform well in poorly textured environments, where point-based odometry is bound to fail.

Yang Gao, Lisa Anne Hendricks, Katherine J. Kuchenbecker et al.

Robots which interact with the physical world will benefit from a fine-grained tactile understanding of objects and surfaces. Additionally, for certain tasks, robots may need to know the haptic properties of an object before touching it. To enable better tactile understanding for robots, we propose a method of classifying surfaces with haptic adjectives (e.g., compressible or smooth) from both visual and physical interaction data. Humans typically combine visual predictions and feedback from physical interactions to accurately predict haptic properties and interact with the world. Inspired by this cognitive pattern, we propose and explore a purely visual haptic prediction model. Purely visual models enable a robot to "feel" without physical interaction. Furthermore, we demonstrate that using both visual and physical interaction signals together yields more accurate haptic classification. Our models take advantage of recent advances in deep neural networks by employing a unified approach to learning features for physical interaction and visual observations. Even though we employ little domain specific knowledge, our model still achieves better results than methods based on hand-designed features.

Yang Gao, Oscar Beijbom, Ning Zhang et al.

Bilinear models has been shown to achieve impressive performance on a wide range of visual tasks, such as semantic segmentation, fine grained recognition and face recognition. However, bilinear features are high dimensional, typically on the order of hundreds of thousands to a few million, which makes them impractical for subsequent analysis. We propose two compact bilinear representations with the same discriminative power as the full bilinear representation but with only a few thousand dimensions. Our compact representations allow back-propagation of classification errors enabling an end-to-end optimization of the visual recognition system. The compact bilinear representations are derived through a novel kernelized analysis of bilinear pooling which provide insights into the discriminative power of bilinear pooling, and a platform for further research in compact pooling methods. Experimentation illustrate the utility of the proposed representations for image classification and few-shot learning across several datasets.