Changhyun Choi

Houjian Yu, Changhyun Choi

Instance segmentation with unseen objects is a challenging problem in unstructured environments. To solve this problem, we propose a robot learning approach to actively interact with novel objects and collect each object's training label for further fine-tuning to improve the segmentation model performance, while avoiding the time-consuming process of manually labeling a dataset. The Singulation-and-Grasping (SaG) policy is trained through end-to-end reinforcement learning. Given a cluttered pile of objects, our approach chooses pushing and grasping motions to break the clutter and conducts object-agnostic grasping for which the SaG policy takes as input the visual observations and imperfect segmentation. We decompose the problem into three subtasks: (1) the object singulation subtask aims to separate the objects from each other, which creates more space that alleviates the difficulty of (2) the collision-free grasping subtask; (3) the mask generation subtask to obtain the self-labeled ground truth masks by using an optical flow-based binary classifier and motion cue post-processing for transfer learning. Our system achieves 70% singulation success rate in simulated cluttered scenes. The interactive segmentation of our system achieves 87.8%, 73.9%, and 69.3% average precision for toy blocks, YCB objects in simulation and real-world novel objects, respectively, which outperforms several baselines.

Alireza Rezazadeh, Athreyi Badithela, Karthik Desingh et al.

Learning multi-object dynamics from visual data using unsupervised techniques is challenging due to the need for robust, object representations that can be learned through robot interactions. This paper presents a novel framework with two new architectures: SlotTransport for discovering object representations from RGB images and SlotGNN for predicting their collective dynamics from RGB images and robot interactions. Our SlotTransport architecture is based on slot attention for unsupervised object discovery and uses a feature transport mechanism to maintain temporal alignment in object-centric representations. This enables the discovery of slots that consistently reflect the composition of multi-object scenes. These slots robustly bind to distinct objects, even under heavy occlusion or absence. Our SlotGNN, a novel unsupervised graph-based dynamics model, predicts the future state of multi-object scenes. SlotGNN learns a graph representation of the scene using the discovered slots from SlotTransport and performs relational and spatial reasoning to predict the future appearance of each slot conditioned on robot actions. We demonstrate the effectiveness of SlotTransport in learning object-centric features that accurately encode both visual and positional information. Further, we highlight the accuracy of SlotGNN in downstream robotic tasks, including challenging multi-object rearrangement and long-horizon prediction. Finally, our unsupervised approach proves effective in the real world. With only minimal additional data, our framework robustly predicts slots and their corresponding dynamics in real-world control tasks.

Youngjin Hong, Houjian Yu, Mingen Li et al.

Learning generalizable policies for robotic manipulation increasingly relies on large-scale models that map language instructions to actions (L2A). However, this one-way paradigm often produces policies that execute tasks without deeper contextual understanding, limiting their ability to generalize or explain their behavior. We argue that the complementary skill of mapping actions back to language (A2L) is essential for developing more holistic grounding. An agent capable of both acting and explaining its actions can form richer internal representations and unlock new paradigms for self-supervised learning. We introduce LACY (Language-Action Cycle), a unified framework that learns such bidirectional mappings within a single vision-language model. LACY is jointly trained on three synergistic tasks: generating parameterized actions from language (L2A), explaining observed actions in language (A2L), and verifying semantic consistency between two language descriptions (L2C). This enables a self-improving cycle that autonomously generates and filters new training data through an active augmentation strategy targeting low-confidence cases, thereby improving the model without additional human labels. Experiments on pick-and-place tasks in both simulation and the real world show that LACY improves task success rates by 56.46% on average and yields more robust language-action grounding for robotic manipulation. Project page: https://vla2026.github.io/LACY/

Yang Yang, Houjian Yu, Xibai Lou et al.

Robotic grasping is one of the most fundamental robotic manipulation tasks and has been the subject of extensive research. However, swiftly teaching a robot to grasp a novel target object in clutter remains challenging. This paper attempts to address the challenge by leveraging object attributes that facilitate recognition, grasping, and rapid adaptation to new domains. In this work, we present an end-to-end encoder-decoder network to learn attribute-based robotic grasping with data-efficient adaptation capability. We first pre-train the end-to-end model with a variety of basic objects to learn generic attribute representation for recognition and grasping. Our approach fuses the embeddings of a workspace image and a query text using a gated-attention mechanism and learns to predict instance grasping affordances. To train the joint embedding space of visual and textual attributes, the robot utilizes object persistence before and after grasping. Our model is self-supervised in a simulation that only uses basic objects of various colors and shapes but generalizes to novel objects in new environments. To further facilitate generalization, we propose two adaptation methods, adversarial adaption and one-grasp adaptation. Adversarial adaptation regulates the image encoder using augmented data of unlabeled images, whereas one-grasp adaptation updates the overall end-to-end model using augmented data from one grasp trial. Both adaptation methods are data-efficient and considerably improve instance grasping performance. Experimental results in both simulation and the real world demonstrate that our approach achieves over 81% instance grasping success rate on unknown objects, which outperforms several baselines by large margins.

Hoseong Jung, Sungil Son, Daesol Cho et al.

Autonomous robotic systems should reason about resource control and its impact on subsequent maneuvers, especially when operating with limited energy budgets or restricted sensing. Learning-based control is effective in handling complex dynamics and represents the problem as a hybrid action space unifying discrete resource usage and continuous maneuvers. However, prior works on hybrid action space have not sufficiently captured the causal dependencies between resource usage and maneuvers. They have also overlooked the multi-modal nature of tactical decisions, both of which are critical in fast-evolving scenarios. In this paper, we propose TART, a Temporal Action Representation learning framework for Tactical resource control and subsequent maneuver generation. TART leverages contrastive learning based on a mutual information objective, designed to capture inherent temporal dependencies in resource-maneuver interactions. These learned representations are quantized into discrete codebook entries that condition the policy, capturing recurring tactical patterns and enabling multi-modal and temporally coherent behaviors. We evaluate TART in two domains where resource deployment is critical: (i) a maze navigation task where a limited budget of discrete actions provides enhanced mobility, and (ii) a high-fidelity air combat simulator in which an F-16 agent operates weapons and defensive systems in coordination with flight maneuvers. Across both domains, TART consistently outperforms hybrid-action baselines, demonstrating its effectiveness in leveraging limited resources and producing context-aware subsequent maneuvers.

Mingen Li, Houjian Yu, Yixuan Huang et al.

Long-horizon routing tasks of deformable linear objects (DLOs), such as cables and ropes, are common in industrial assembly lines and everyday life. These tasks are particularly challenging because they require robots to manipulate DLO with long-horizon planning and reliable skill execution. Successfully completing such tasks demands adapting to their nonlinear dynamics, decomposing abstract routing goals, and generating multi-step plans composed of multiple skills, all of which require accurate high-level reasoning during execution. In this paper, we propose a fully autonomous hierarchical framework for solving challenging DLO routing tasks. Given an implicit or explicit routing goal expressed in language, our framework leverages vision-language models~(VLMs) for in-context high-level reasoning to synthesize feasible plans, which are then executed by low-level skills trained via reinforcement learning. To improve robustness in long horizons, we further introduce a failure recovery mechanism that reorients the DLO into insertion-feasible states. Our approach generalizes to diverse scenes involving object attributes, spatial descriptions, as well as implicit language commands. It outperforms the next best baseline method by nearly 50% and achieves an overall success rate of 92.5% across long-horizon routing scenarios.

Changhyun Choi, Sungha Kim, H. Jin Kim

Recently, it has been shown that investing computing resources in searching for good initial noise for a text-to-image diffusion model helps improve performance. However, previous studies required external models to evaluate the resulting images, which is impossible on GPUs with small VRAM. For these reasons, we apply Best-of-N inference-time scaling to algorithms that optimize the initial noise of a diffusion model without external models across multiple datasets and backbones. We demonstrate that inference-time scaling for text-to-image diffusion models in this setting quickly reaches a performance plateau, and a relatively small number of optimization steps suffices to achieve the maximum achievable performance with each algorithm.

Alireza Rezazadeh, Changhyun Choi

Object-centric representation is an essential abstraction for forward prediction. Most existing forward models learn this representation through extensive supervision (e.g., object class and bounding box) although such ground-truth information is not readily accessible in reality. To address this, we introduce KINet (Keypoint Interaction Network) -- an end-to-end unsupervised framework to reason about object interactions based on a keypoint representation. Using visual observations, our model learns to associate objects with keypoint coordinates and discovers a graph representation of the system as a set of keypoint embeddings and their relations. It then learns an action-conditioned forward model using contrastive estimation to predict future keypoint states. By learning to perform physical reasoning in the keypoint space, our model automatically generalizes to scenarios with a different number of objects, novel backgrounds, and unseen object geometries. Experiments demonstrate the effectiveness of our model in accurately performing forward prediction and learning plannable object-centric representations for downstream robotic pushing manipulation tasks.

Yang Yang, Yuanhao Liu, Hengyue Liang et al.

Robotic grasping is one of the most fundamental robotic manipulation tasks and has been actively studied. However, how to quickly teach a robot to grasp a novel target object in clutter remains challenging. This paper attempts to tackle the challenge by leveraging object attributes that facilitate recognition, grasping, and quick adaptation. In this work, we introduce an end-to-end learning method of attribute-based robotic grasping with one-grasp adaptation capability. Our approach fuses the embeddings of a workspace image and a query text using a gated-attention mechanism and learns to predict instance grasping affordances. Besides, we utilize object persistence before and after grasping to learn a joint metric space of visual and textual attributes. Our model is self-supervised in a simulation that only uses basic objects of various colors and shapes but generalizes to novel objects and real-world scenes. We further demonstrate that our model is capable of adapting to novel objects with only one grasp data and improving instance grasping performance significantly. Experimental results in both simulation and the real world demonstrate that our approach achieves over 80\% instance grasping success rate on unknown objects, which outperforms several baselines by large margins.

Xibai Lou, Yang Yang, Changhyun Choi

Grasping a novel target object in constrained environments (e.g., walls, bins, and shelves) requires intensive reasoning about grasp pose reachability to avoid collisions with the surrounding structures. Typical 6-DoF robotic grasping systems rely on the prior knowledge about the environment and intensive planning computation, which is ungeneralizable and inefficient. In contrast, we propose a novel Collision-Aware Reachability Predictor (CARP) for 6-DoF grasping systems. The CARP learns to estimate the collision-free probabilities for grasp poses and significantly improves grasping in challenging environments. The deep neural networks in our approach are trained fully by self-supervision in simulation. The experiments in both simulation and the real world show that our approach achieves more than 75% grasping rate on novel objects in various surrounding structures. The ablation study demonstrates the effectiveness of the CARP, which improves the 6-DoF grasping rate by 95.7%.

Xibai Lou, Yang Yang, Changhyun Choi

Motivated by the stringent requirements of unstructured real-world where a plethora of unknown objects reside in arbitrary locations of the surface, we propose a voxel-based deep 3D Convolutional Neural Network (3D CNN) that generates feasible 6-DoF grasp poses in unrestricted workspace with reachability awareness. Unlike the majority of works that predict if a proposed grasp pose within the restricted workspace will be successful solely based on grasp pose stability, our approach further learns a reachability predictor that evaluates if the grasp pose is reachable or not from robot's own experience. To avoid the laborious real training data collection, we exploit the power of simulation to train our networks on a large-scale synthetic dataset. This work is an early attempt that simultaneously evaluates grasping reachability from learned knowledge while proposing feasible grasp poses with 3D CNN. Experimental results in both simulation and real-world demonstrate that our approach outperforms several other methods and achieves 82.5% grasping success rate on unknown objects.

Hengyue Liang, Xibai Lou, Yang Yang et al.

This paper introduces a challenging object grasping task and proposes a self-supervised learning approach. The goal of the task is to grasp an object which is not feasible with a single parallel gripper, but only with harnessing environment fixtures (e.g., walls, furniture, heavy objects). This Slide-to-Wall grasping task assumes no prior knowledge except the partial observation of a target object. Hence the robot should learn an effective policy given a scene observation that may include the target object, environmental fixtures, and any other disturbing objects. We formulate the problem as visual affordances learning for which Target-Oriented Deep Q-Network (TO-DQN) is proposed to efficiently learn visual affordance maps (i.e., Q-maps) to guide robot actions. Since the training necessitates robot's exploration and collision with the fixtures, TO-DQN is first trained safely with a simulated robot manipulator and then applied to a real robot. We empirically show that TO-DQN can learn to solve the task in different environment settings in simulation and outperforms a standard and a variant of Deep Q-Network (DQN) in terms of training efficiency and robustness. The testing performance in both simulation and real-robot experiments shows that the policy trained by TO-DQN achieves comparable performance to humans.

Yang Yang, Hengyue Liang, Changhyun Choi

We study an emerging problem named "grasping the invisible" in robotic manipulation, in which a robot is tasked to grasp an initially invisible target object via a sequence of pushing and grasping actions. In this problem, pushes are needed to search for the target and rearrange cluttered objects around it to enable effective grasps. We propose to solve the problem by formulating a deep learning approach in a critic-policy format. The target-oriented motion critic, which maps both visual observations and target information to the expected future rewards of pushing and grasping motion primitives, is learned via deep Q-learning. We divide the problem into two subtasks, and two policies are proposed to tackle each of them, by combining the critic predictions and relevant domain knowledge. A Bayesian-based policy accounting for past action experience performs pushing to search for the target; once the target is found, a classifier-based policy coordinates target-oriented pushing and grasping to grasp the target in clutter. The motion critic and the classifier are trained in a self-supervised manner through robot-environment interactions. Our system achieves a 93% and 87% task success rate on each of the two subtasks in simulation and an 85% task success rate in real robot experiments on the whole problem, which outperforms several baselines by large margins. Supplementary material is available at https://sites.google.com/umn.edu/grasping-invisible.