Condensed datasets offer a compact representation of larger datasets, but training models directly on them or using them to enhance model performance through knowledge distillation (KD) can result in suboptimal outcomes due to limited information. To address this, we propose a method that expands condensed datasets using model inversion, a technique for generating synthetic data based on the impressions of a pre-trained model on its training data. This approach is particularly well-suited for KD scenarios, as the teacher model is already pre-trained and retains knowledge of the original training data. By creating synthetic data that complements the condensed samples, we enrich the training set and better approximate the underlying data distribution, leading to improvements in student model accuracy during knowledge distillation. Our method demonstrates significant gains in KD accuracy compared to using condensed datasets alone and outperforms standard model inversion-based KD methods by up to 11.4% across various datasets and model architectures. Importantly, it remains effective even when using as few as one condensed sample per class, and can also enhance performance in few-shot scenarios where only limited real data samples are available.

Text-to-image diffusion models often use a fixed number of denoising steps, balancing time costs and image quality. However, the optimal number of steps depends on the complexity of the input text prompt. We propose an adaptive diffusion controller that dynamically adjusts the number of steps to generate high-quality images efficiently, without additional model training. By leveraging a mixture of step schedules with varying step sizes and evaluating the error term discrepancy at each timestep, our method transitions between schedules to optimize performance. Experiments on COCO and DiffusionDB show that our approach reduces inference time while maintaining visual fidelity, offering a more efficient alternative for text-to-image diffusion models.

Sampling-based motion planning under task constraints is challenging because the null-measure constraint manifold in the configuration space makes rejection sampling extremely inefficient, if not impossible. This paper presents a learning-based sampling strategy for constrained motion planning problems. We investigate the use of two well-known deep generative models, the Conditional Variational Autoencoder (CVAE) and the Conditional Generative Adversarial Net (CGAN), to generate constraint-satisfying sample configurations. Instead of precomputed graphs, we use generative models conditioned on constraint parameters for approximating the constraint manifold. This approach allows for the efficient drawing of constraint-satisfying samples online without any need for modification of available sampling-based motion planning algorithms. We evaluate the efficiency of these two generative models in terms of their sampling accuracy and coverage of sampling distribution. Simulations and experiments are also conducted for different constraint tasks on two robotic platforms.

This paper introduces a general state estimation framework fusing multiple sensor information for hybrid wheeled-legged robots performing mobile manipulation tasks. At the core of the state estimator is a novel unified odometry for hybrid locomotion which can seamlessly maintain tracking and has no need to switch between stepping and rolling modes. To the best of our knowledge, the proposed odometry is the first work in this area. It is calculated based on the robot kinematics and instantaneous contact points of wheels with sensor inputs from IMU, joint encoders, joint torque sensors estimating wheel contact status, as well as RGB-D camera detecting geometric features of the terrain (e.g. elevation and surface normal vector). Subsequently, the odometry output is utilized as the motion model of a 3D Lidar map-based Monte Carlo Localization module for drift-free state estimation. As part of the framework, visual localization is integrated to provide high precision guidance for the robot movement relative to an object of interest. The proposed approach was verified thoroughly by two experiments conducted on the Pholus robot with OptiTrack measurements as ground truth.

This paper proposes an improved supervised autonomy framework for remote teleoperation of a quadrupedal bimanual mobile manipulator in an unknown environment, with the usage of advanced perception technology and allowing the operator to easily assist the robot with decision making for executing tasks on the fly. First, the perception system uses lightweight deep neural network-based Single Shot Detector (SSD) MobileNet on RGB images to detect objects and highlight them to the human operator via an intuitive interactive visualization interface. After object and action selections are made by the operator, segmentation of object point cloud and 3D surfaces based on random sample consensus is performed, followed by object pose localization by using keypoint extraction. Based on the localized object, mobile manipulation motion to perform the operator-selected action is planned and executed with the help of a state estimator for the hybrid wheel-legged robot. Thanks to the autonomy of the robot in perception and manipulation, the complexity of teleoperating the robot is reduced to specifying the essential task objectives. Experimental results on the real robot, with full system integration, for 2 task scenarios, namely passage clearing and object retrieval, demonstrate a high average success rate of 92.2% over a total of 90 trials.

The domain of robotics is challenging to apply deep reinforcement learning due to the need for large amounts of data and for ensuring safety during learning. Curriculum learning has shown good performance in terms of sample-efficient deep learning. In this paper, we propose an algorithm (named GloCAL) that creates a curriculum for an agent to learn multiple discrete tasks, based on clustering tasks according to their evaluation scores. From the highest-performing cluster, a global task representative of the cluster is identified for learning a global policy that transfers to subsequently formed new clusters, while remaining tasks in the cluster are learnt as local policies. The efficacy and efficiency of our GloCAL algorithm are compared with other approaches in the domain of grasp learning for 49 objects with varied object complexity and grasp difficulty from the EGAD! dataset. The results show that GloCAL is able to learn to grasp 100% of the objects, whereas other approaches achieve at most 86% despite being given 1.5× longer training time.

6D object pose estimation is widely applied in robotic tasks such as grasping and manipulation. Prior methods using RGB-only images are vulnerable to heavy occlusion and poor illumination, so it is important to complement them with depth information. However, existing methods using RGB-D data cannot adequately exploit consistent and complementary information between RGB and depth modalities. In this paper, we present a novel method to effectively consider the correlation within and across both modalities with attention mechanism to learn discriminative and compact multi-modal features. Then, effective fusion strategies for intra- and inter-correlation modules are explored to ensure efficient information flow between RGB and depth. To our best knowledge, this is the first work to explore effective intra- and inter-modality fusion in 6D pose estimation. The experimental results show that our method can achieve the state-of-the-art performance on LineMOD and YCB-Video dataset. We also demonstrate that the proposed method can benefit a real-world robot grasping task by providing accurate object pose estimation.