We built a differentiable ray tracer for meshes and a novel multi-stage constrained based optimization pipeline to reconstruct unseen scenes without training data using a single observation. Furthermore, we validated the accuracy of our scene reconstructions by applying our method to a zero-shot robot grasping task.

In this paper we argue for the following points in machine and robot learning:
▶ Learning algorithms shouldn’t ignore the most accurate predictive models of our world.
▶ Performance increases with more data, yet it remains insufficient for robot autonomy.
▶ We cannot trust the probabilities of a model that we do not understand.

Current few-shot learning models have billions of parameters and use billions of training samples. In this paper we present a few-shot classifier that:
▶ Generalizes with only few training samples.
▶ Uses only tenths of parameters.
The proposed model combines probabilistic reasoning, physics-based rendering, and a neural similarity metric to achieve strong few-shot generalization with limited data and parameters.

We present an unsupervised method for calculating a difficulty sample score. Our method:
▶ Does not require any modification to the model.
▶ Does not require external supervision.
▶ It can be easily applied to a wide range of machine learning tasks.
We provide results for image classification, image segmentation, and object detection tasks.

In this paper, we introduce PhysWM, a framework for robot learning that unifies differentiable physics, rendering, and Bayesian inference into an adaptable physical world model. The result is a more efficient and interpretable approach that allows robots to estimate uncertainty, improve their internal models through exploration.

We extend saliency-method sanity checks to object detectors and show that detector architecture often matters more than the explanation method itself.

We present an open-source toolkit for explaining object detector predictions, including both class decisions and bounding boxes, across multiple detector architectures.

We formulate robot dance generation as trajectory optimization aligned with musical structure, enabling expert-guided, imitative, and automatically generated choreographies.

We combine context recognition modules with a two-layer Bayesian network to infer astronaut intentions for context-adaptive robot assistance in space missions.

We optimize CNN architectures and training hyperparameters for automatic mosquito classification, improving species recognition and Aedes detection from entomological images.

We introduce PAZ, a hierarchical perception library for building reusable training and inference pipelines across robotics and computer-vision tasks.

We introduce the action score as a lightweight callback-based measure of sample difficulty that helps analyze model behavior and dataset bias without extra supervision.

We use black-box optimization to tune detector scales and input sizes, improving object-detection performance without manual hyperparameter search.

We proposed a general framework for designing real-time CNNs for robotic platforms. By eliminating fully connected layers and leveraging depth-wise separable convolutions, we reduced the model parameters by 80% while maintaining human-level accuracy (96% for gender, 66% for emotion).

We learn multiple context-dependent underwater vehicle dynamics with recurrent models, improving robustness to changing operating conditions.

We introduce a dataset and deep model for classifying and describing dangerous scenes such as fires, injuries, and accidents from a single image.

This technical report surveys scene understanding for robotics and introduces early work on anomaly description and classification with deep learning.

We study attitude synchronization for dumbbell satellites using a generalized Hamiltonian formulation and observer-based control design.