Unity RL Playground: A Versatile Reinforcement Learning Framework for Mobile Robots

This paper introduces Unity RL Playground, an open-source reinforcement learning framework built on top of Unity ML-Agents. Unity RL Playground automates the process of training mobile robots to perform various locomotion tasks such as walking, running, and jumping in simulation, with the potential for seamless transfer to real hardware. Key features include one-click training for imported robot models, universal compatibility with diverse robot configurations, multi-mode motion learning capabilities, and extreme performance testing to aid in robot design optimization and morphological evolution. The attached video can be found at https://linqi-ye.github.io/video/iros25.mp4 and the code is coming soon.

NeuGrasp: Generalizable Neural Surface Reconstruction with Background Priors for Material-Agnostic Object Grasp Detection

Robotic grasping in scenes with transparent and specular objects presents great challenges for methods relying on accurate depth information. In this paper, we introduce NeuGrasp, a neural surface reconstruction method that leverages background priors for material-agnostic grasp detection. NeuGrasp integrates transformers and global prior volumes to aggregate multi-view features with spatial encoding, enabling robust surface reconstruction in narrow and sparse viewing conditions. By focusing on foreground objects through residual feature enhancement and refining spatial perception with an occupancy-prior volume, NeuGrasp excels in handling objects with transparent and specular surfaces. Extensive experiments in both simulated and real-world scenarios show that NeuGrasp outperforms state-of-the-art methods in grasping while maintaining comparable reconstruction quality. More details are available at https://neugrasp.github.io/.

PEARL: Towards Permutation-Resilient LLMs

The in-context learning (ICL) capability of large language models (LLMs) enables them to perform challenging tasks using provided demonstrations. However, ICL is highly sensitive to the ordering of demonstrations, leading to instability in predictions. This paper shows that this vulnerability can be exploited to design a natural attack - difficult for model providers to detect - that achieves nearly 80% success rate on LLaMA-3 by simply permuting the demonstrations. Existing mitigation methods primarily rely on post-processing and fail to enhance the model's inherent robustness to input permutations, raising concerns about safety and reliability of LLMs. To address this issue, we propose Permutation-resilient learning (PEARL), a novel framework based on distributionally robust optimization (DRO), which optimizes model performance against the worst-case input permutation. Specifically, PEARL consists of a permutation-proposal network (P-Net) and the LLM. The P-Net generates the most challenging permutations by treating it as an optimal transport problem, which is solved using an entropy-constrained Sinkhorn algorithm. Through minimax optimization, the P-Net and the LLM iteratively optimize against each other, progressively improving the LLM's robustness. Experiments on synthetic pre-training and real-world instruction tuning tasks demonstrate that PEARL effectively mitigates permutation attacks and enhances performance. Notably, despite being trained on fewer shots and shorter contexts, PEARL achieves performance gains of up to 40% when scaled to many-shot and long-context scenarios, highlighting its efficiency and generalization capabilities.

Señorita-2M: A High-Quality Instruction-based Dataset for General Video Editing by Video Specialists

Recent advancements in video generation have spurred the development of video editing techniques, which can be divided into inversion-based and end-to-end methods. However, current video editing methods still suffer from several challenges. Inversion-based methods, though training-free and flexible, are time-consuming during inference, struggle with fine-grained editing instructions, and produce artifacts and jitter. On the other hand, end-to-end methods, which rely on edited video pairs for training, offer faster inference speeds but often produce poor editing results due to a lack of high-quality training video pairs. In this paper, to close the gap in end-to-end methods, we introduce Se\~norita-2M, a high-quality video editing dataset. Se\~norita-2M consists of approximately 2 millions of video editing pairs. It is built by crafting four high-quality, specialized video editing models, each crafted and trained by our team to achieve state-of-the-art editing results. We also propose a filtering pipeline to eliminate poorly edited video pairs. Furthermore, we explore common video editing architectures to identify the most effective structure based on current pre-trained generative model. Extensive experiments show that our dataset can help to yield remarkably high-quality video editing results. More details are available at https://senorita.github.io.

Episodic Novelty Through Temporal Distance

Exploration in sparse reward environments remains a significant challenge in reinforcement learning, particularly in Contextual Markov Decision Processes (CMDPs), where environments differ across episodes. Existing episodic intrinsic motivation methods for CMDPs primarily rely on count-based approaches, which are ineffective in large state spaces, or on similarity-based methods that lack appropriate metrics for state comparison. To address these shortcomings, we propose Episodic Novelty Through Temporal Distance (ETD), a novel approach that introduces temporal distance as a robust metric for state similarity and intrinsic reward computation. By employing contrastive learning, ETD accurately estimates temporal distances and derives intrinsic rewards based on the novelty of states within the current episode. Extensive experiments on various benchmark tasks demonstrate that ETD significantly outperforms state-of-the-art methods, highlighting its effectiveness in enhancing exploration in sparse reward CMDPs.

DeepMF: Deep Motion Factorization for Closed-Loop Safety-Critical Driving Scenario Simulation

Safety-critical traffic scenarios are of great practical relevance to evaluating the robustness of autonomous driving (AD) systems. Given that these long-tail events are extremely rare in real-world traffic data, there is a growing body of work dedicated to the automatic traffic scenario generation. However, nearly all existing algorithms for generating safety-critical scenarios rely on snippets of previously recorded traffic events, transforming normal traffic flow into accident-prone situations directly. In other words, safety-critical traffic scenario generation is hindsight and not applicable to newly encountered and open-ended traffic events.In this paper, we propose the Deep Motion Factorization (DeepMF) framework, which extends static safety-critical driving scenario generation to closed-loop and interactive adversarial traffic simulation. DeepMF casts safety-critical traffic simulation as a Bayesian factorization that includes the assignment of hazardous traffic participants, the motion prediction of selected opponents, the reaction estimation of autonomous vehicle (AV) and the probability estimation of the accident occur. All the aforementioned terms are calculated using decoupled deep neural networks, with inputs limited to the current observation and historical states. Consequently, DeepMF can effectively and efficiently simulate safety-critical traffic scenarios at any triggered time and for any duration by maximizing the compounded posterior probability of traffic risk. Extensive experiments demonstrate that DeepMF excels in terms of risk management, flexibility, and diversity, showcasing outstanding performance in simulating a wide range of realistic, high-risk traffic scenarios.

Correcting Large Language Model Behavior via Influence Function

Recent advancements in AI alignment techniques have significantly improved the alignment of large language models (LLMs) with static human preferences. However, the dynamic nature of human preferences can render some prior training data outdated or even erroneous, ultimately causing LLMs to deviate from contemporary human preferences and societal norms. Existing methodologies, whether they involve the curation of new data for continual alignment or the manual correction of outdated data for re-alignment, demand costly human resources. To address this challenge, we propose a novel approach, Large Language Model Behavior Correction with Influence Function Recall and Post-Training (LANCET), which requires no human involvement. LANCET consists of two phases: (1) using influence functions to identify the training data that significantly impact undesirable model outputs, and (2) applying an Influence function-driven Bregman Optimization (IBO) technique to adjust the model's behavior based on these influence distributions. Our experiments demonstrate that LANCET effectively and efficiently correct inappropriate behaviors of LLMs. Furthermore, LANCET can outperform methods that rely on collecting human preferences, and it enhances the interpretability of learning human preferences within LLMs.

Can GNNs Learn Link Heuristics? A Concise Review and Evaluation of Link Prediction Methods

This paper explores the ability of Graph Neural Networks (GNNs) in learning various forms of information for link prediction, alongside a brief review of existing link prediction methods. Our analysis reveals that GNNs cannot effectively learn structural information related to the number of common neighbors between two nodes, primarily due to the nature of set-based pooling of the neighborhood aggregation scheme. Also, our extensive experiments indicate that trainable node embeddings can improve the performance of GNN-based link prediction models. Importantly, we observe that the denser the graph, the greater such the improvement. We attribute this to the characteristics of node embeddings, where the link state of each link sample could be encoded into the embeddings of nodes that are involved in the neighborhood aggregation of the two nodes in that link sample. In denser graphs, every node could have more opportunities to attend the neighborhood aggregation of other nodes and encode states of more link samples to its embedding, thus learning better node embeddings for link prediction. Lastly, we demonstrate that the insights gained from our research carry important implications in identifying the limitations of existing link prediction methods, which could guide the future development of more robust algorithms.

Efficient Collaborative Navigation through Perception Fusion for Multi-Robots in Unknown Environments

For tasks conducted in unknown environments with efficiency requirements, real-time navigation of multi-robot systems remains challenging due to unfamiliarity with surroundings.In this paper, we propose a novel multi-robot collaborative planning method that leverages the perception of different robots to intelligently select search directions and improve planning efficiency. Specifically, a foundational planner is employed to ensure reliable exploration towards targets in unknown environments and we introduce Graph Attention Architecture with Information Gain Weight(GIWT) to synthesizes the information from the target robot and its teammates to facilitate effective navigation around obstacles.In GIWT, after regionally encoding the relative positions of the robots along with their perceptual features, we compute the shared attention scores and incorporate the information gain obtained from neighboring robots as a supplementary weight. We design a corresponding expert data generation scheme to simulate real-world decision-making conditions for network training. Simulation experiments and real robot tests demonstrates that the proposed method significantly improves efficiency and enables collaborative planning for multiple robots. Our method achieves approximately 82% accuracy on the expert dataset and reduces the average path length by about 8% and 6% across two types of tasks compared to the fundamental planner in ROS tests, and a path length reduction of over 6% in real-world experiments.

Equilibrium Adaptation-Based Control for Track Stand of Single-Track Two-Wheeled Robots

Stationary balance control is challenging for single-track two-wheeled (STTW) robots due to the lack of elegant balancing mechanisms and the conflict between the limited attraction domain and external disturbances. To address the absence of balancing mechanisms, we draw inspiration from cyclists and leverage the track stand maneuver, which relies solely on steering and rear-wheel actuation. To achieve accurate tracking in the presence of matched and mismatched disturbances, we propose an equilibrium adaptation-based control (EABC) scheme that can be seamlessly integrated with standard disturbance observers and controllers. This scheme enables adaptation to slow-varying disturbances by utilizing a disturbed equilibrium estimator, effectively handling both matched and mismatched disturbances in a unified manner while ensuring accurate tracking with zero steady-state error. We integrate the EABC scheme with nonlinear model predictive control (MPC) for the track stand of STTW robots and validate its effectiveness through two experimental scenarios. Our method demonstrates significant improvements in tracking accuracy, reducing errors by several orders of magnitude.