Michael Wan

Sai Kumar Reddy Manne, Shaotong Zhu, Sarah Ostadabbas et al.

Respiration is a critical vital sign for infants, and continuous respiratory monitoring is particularly important for newborns. However, neonates are sensitive and contact-based sensors present challenges in comfort, hygiene, and skin health, especially for preterm babies. As a step toward fully automatic, continuous, and contactless respiratory monitoring, we develop a deep-learning method for estimating respiratory rate and waveform from plain video footage in natural settings. Our automated infant respiration flow-based network (AIRFlowNet) combines video-extracted optical flow input and spatiotemporal convolutional processing tuned to the infant domain. We support our model with the first public annotated infant respiration dataset with 125 videos (AIR-125), drawn from eight infant subjects, set varied pose, lighting, and camera conditions. We include manual respiration annotations and optimize AIRFlowNet training on them using a novel spectral bandpass loss function. When trained and tested on the AIR-125 infant data, our method significantly outperforms other state-of-the-art methods in respiratory rate estimation, achieving a mean absolute error of $\sim$2.9 breaths per minute, compared to $\sim$4.7--6.2 for other public models designed for adult subjects and more uniform environments.

Sai Kumar Reddy Manne, Michael Wan

Automatic white balancing (AWB), one of the first steps in an integrated signal processing (ISP) pipeline, aims to correct the color cast induced by the scene illuminant. An incorrect white balance (WB) setting or AWB failure can lead to an undesired blue or red tint in the rendered sRGB image. To address this, recent methods pose the post-capture WB correction problem as an image-to-image translation task and train deep neural networks to learn the necessary color adjustments at a lower resolution. These low resolution outputs are post-processed to generate high resolution WB corrected images, forming a bottleneck in the end-to-end run time. In this paper we present a 3D Lookup Table (LUT) based WB correction model called WB LUTs that can generate high resolution outputs in real time. We introduce a contrastive learning framework with a novel hard sample mining strategy, which improves the WB correction quality of baseline 3D LUTs by 25.5%. Experimental results demonstrate that the proposed WB LUTs perform competitively against state-of-the-art models on two benchmark datasets while being 300 times faster using 12.7 times less memory. Our model and code are available at https://github.com/skrmanne/3DLUT_sRGB_WB.

Yu Zhang, Yuqi Xie, Huihan Liu et al.

Imitation learning advances robot capabilities by enabling the acquisition of diverse behaviors from human demonstrations. However, large-scale datasets used for policy training often introduce substantial variability in quality, which can negatively impact performance. As a result, automatically curating datasets by filtering low-quality samples to improve quality becomes essential. Existing robotic curation approaches rely on costly manual annotations and perform curation at a coarse granularity, such as the dataset or trajectory level, failing to account for the quality of individual state-action pairs. To address this, we introduce SCIZOR, a self-supervised data curation framework that filters out low-quality state-action pairs to improve the performance of imitation learning policies. SCIZOR targets two complementary sources of low-quality data: suboptimal data, which hinders learning with undesirable actions, and redundant data, which dilutes training with repetitive patterns. SCIZOR leverages a self-supervised task progress predictor for suboptimal data to remove samples lacking task progression, and a deduplication module operating on joint state-action representation for samples with redundant patterns. Empirically, we show that SCIZOR enables imitation learning policies to achieve higher performance with less data, yielding an average improvement of 15.4% across multiple benchmarks. More information is available at: https://ut-austin-rpl.github.io/SCIZOR/

Michael Wan, Shaotong Zhu, Lingfei Luan et al.

We lay the groundwork for research in the algorithmic comprehension of infant faces, in anticipation of applications from healthcare to psychology, especially in the early prediction of developmental disorders. Specifically, we introduce the first-ever dataset of infant faces annotated with facial landmark coordinates and pose attributes, demonstrate the inadequacies of existing facial landmark estimation algorithms in the infant domain, and train new state-of-the-art models that significantly improve upon those algorithms using domain adaptation techniques. We touch on the closely related task of facial detection for infants, and also on a challenging case study of infrared baby monitor images gathered by our lab as part of in-field research into the aforementioned developmental issues.

Michael Wan, Jian Peng, Tanmay Gangwani

Meta-reinforcement learning (meta-RL) algorithms allow for agents to learn new behaviors from small amounts of experience, mitigating the sample inefficiency problem in RL. However, while meta-RL agents can adapt quickly to new tasks at test time after experiencing only a few trajectories, the meta-training process is still sample-inefficient. Prior works have found that in the multi-task RL setting, relabeling past transitions and thus sharing experience among tasks can improve sample efficiency and asymptotic performance. We apply this idea to the meta-RL setting and devise a new relabeling method called Hindsight Foresight Relabeling (HFR). We construct a relabeling distribution using the combination of "hindsight", which is used to relabel trajectories using reward functions from the training task distribution, and "foresight", which takes the relabeled trajectories and computes the utility of each trajectory for each task. HFR is easy to implement and readily compatible with existing meta-RL algorithms. We find that HFR improves performance when compared to other relabeling methods on a variety of meta-RL tasks.

Michael Wan, Tanmay Gangwani, Jian Peng

Deep reinforcement learning (RL) algorithms have achieved great success on a wide variety of sequential decision-making tasks. However, many of these algorithms suffer from high sample complexity when learning from scratch using environmental rewards, due to issues such as credit-assignment and high-variance gradients, among others. Transfer learning, in which knowledge gained on a source task is applied to more efficiently learn a different but related target task, is a promising approach to improve the sample complexity in RL. Prior work has considered using pre-trained teacher policies to enhance the learning of the student policy, albeit with the constraint that the teacher and the student MDPs share the state-space or the action-space. In this paper, we propose a new framework for transfer learning where the teacher and the student can have arbitrarily different state- and action-spaces. To handle this mismatch, we produce embeddings which can systematically extract knowledge from the teacher policy and value networks, and blend it into the student networks. To train the embeddings, we use a task-aligned loss and show that the representations could be enriched further by adding a mutual information loss. Using a set of challenging simulated robotic locomotion tasks involving many-legged centipedes, we demonstrate successful transfer learning in situations when the teacher and student have different state- and action-spaces.