Harry Zhang

Ben Eisner, Harry Zhang, David Held · cmu

We explore a novel method to perceive and manipulate 3D articulated objects that generalizes to enable a robot to articulate unseen classes of objects. We propose a vision-based system that learns to predict the potential motions of the parts of a variety of articulated objects to guide downstream motion planning of the system to articulate the objects. To predict the object motions, we train a neural network to output a dense vector field representing the point-wise motion direction of the points in the point cloud under articulation. We then deploy an analytical motion planner based on this vector field to achieve a policy that yields maximum articulation. We train the vision system entirely in simulation, and we demonstrate the capability of our system to generalize to unseen object instances and novel categories in both simulation and the real world, deploying our policy on a Sawyer robot with no finetuning. Results show that our system achieves state-of-the-art performance in both simulated and real-world experiments.

Jingquan Wang, Andrew Negrut, Hongyu Wang et al.

We introduce SimBench, a benchmark designed to evaluate the proficiency of simulator-oriented LLMs (S-LLMs) in generating digital twins (DTs) that can be used in simulators for virtual testing. Given a collection of S-LLMs, this benchmark ranks them according to their ability to produce high-quality DTs. We demonstrate this by comparing over 33 open- and closed-source S-LLMs. Using multi-turn interactions, SimBench employs an LLM-as-a-judge (J-LLM) that leverages both predefined rules and human-in-the-loop guidance to assign scores for the DTs generated by the S-LLM, thus providing a consistent and expert-inspired evaluation protocol. The J-LLM is specific to a simulator, and herein the proposed benchmarking approach is demonstrated in conjunction with the open-sourceChrono multi-physics simulator. Chrono provided the backdrop used to assess an S-LLM in relation to the latter's ability to create digital twins for multibody dynamics, finite element analysis, vehicle dynamics, robotic dynamics, and sensor simulations. The proposed benchmarking principle is broadly applicable and enables the assessment of an S-LLM's ability to generate digital twins for other simulation packages, e.g., ANSYS, ABAQUS, OpenFOAM, StarCCM+, IsaacSim, and pyBullet.

Yupu Yao, Shangqi Deng, Zihan Cao et al.

Diffusion models have exhibited promising progress in video generation. However, they often struggle to retain consistent details within local regions across frames. One underlying cause is that traditional diffusion models approximate Gaussian noise distribution by utilizing predictive noise, without fully accounting for the impact of inherent information within the input itself. Additionally, these models emphasize the distinction between predictions and references, neglecting information intrinsic to the videos. To address this limitation, inspired by the self-attention mechanism, we propose a novel text-to-video (T2V) generation network structure based on diffusion models, dubbed Additional Perturbation for Latent noise with Adversarial training (APLA). Our approach only necessitates a single video as input and builds upon pre-trained stable diffusion networks. Notably, we introduce an additional compact network, known as the Video Generation Transformer (VGT). This auxiliary component is designed to extract perturbations from the inherent information contained within the input, thereby refining inconsistent pixels during temporal predictions. We leverage a hybrid architecture of transformers and convolutions to compensate for temporal intricacies, enhancing consistency between different frames within the video. Experiments demonstrate a noticeable improvement in the consistency of the generated videos both qualitatively and quantitatively.

Chuer Pan, Brian Okorn, Harry Zhang et al.

How do we imbue robots with the ability to efficiently manipulate unseen objects and transfer relevant skills based on demonstrations? End-to-end learning methods often fail to generalize to novel objects or unseen configurations. Instead, we focus on the task-specific pose relationship between relevant parts of interacting objects. We conjecture that this relationship is a generalizable notion of a manipulation task that can transfer to new objects in the same category; examples include the relationship between the pose of a pan relative to an oven or the pose of a mug relative to a mug rack. We call this task-specific pose relationship "cross-pose" and provide a mathematical definition of this concept. We propose a vision-based system that learns to estimate the cross-pose between two objects for a given manipulation task using learned cross-object correspondences. The estimated cross-pose is then used to guide a downstream motion planner to manipulate the objects into the desired pose relationship (placing a pan into the oven or the mug onto the mug rack). We demonstrate our method's capability to generalize to unseen objects, in some cases after training on only 10 demonstrations in the real world. Results show that our system achieves state-of-the-art performance in both simulated and real-world experiments across a number of tasks. Supplementary information and videos can be found at https://sites.google.com/view/tax-pose/home.

Lillian Zhou, Yuxin Ding, Mingqing Chen et al.

Automatic speech recognition (ASR) models are typically trained on large datasets of transcribed speech. As language evolves and new terms come into use, these models can become outdated and stale. In the context of models trained on the server but deployed on edge devices, errors may result from the mismatch between server training data and actual on-device usage. In this work, we seek to continually learn from on-device user corrections through Federated Learning (FL) to address this issue. We explore techniques to target fresh terms that the model has not previously encountered, learn long-tail words, and mitigate catastrophic forgetting. In experimental evaluations, we find that the proposed techniques improve model recognition of fresh terms, while preserving quality on the overall language distribution.

Bo-Hsun Chen, Peter Negrut, Thomas Liang et al.

NASA's POLAR dataset contains approximately 2,600 pairs of high dynamic range stereo photos captured across 13 varied terrain scenarios, including areas with sparse or dense rock distributions, craters, and rocks of different sizes. The purpose of these photos is to spur development in robotics, AI-based perception, and autonomous navigation. Acknowledging a scarcity of lunar images from around the lunar poles, NASA Ames produced on Earth but in controlled conditions images that resemble rover operating conditions from these regions of the Moon. We report on the outcomes of an effort aimed at accomplishing two tasks. In Task 1, we provided bounding boxes and semantic segmentation information for all the images in NASA's POLAR dataset. This effort resulted in 23,000 labels and semantic segmentation annotations pertaining to rocks, shadows, and craters. In Task 2, we generated the digital twins of the 13 scenarios that have been used to produce all the photos in the POLAR dataset. Specifically, for each of these scenarios, we produced individual meshes, texture information, and material properties associated with the ground and the rocks in each scenario. This allows anyone with a camera model to synthesize images associated with any of the 13 scenarios of the POLAR dataset. Effectively, one can generate as many semantically labeled synthetic images as desired -- with different locations and exposure values in the scene, for different positions of the sun, with or without the presence of active illumination, etc. The benefit of this work is twofold. Using outcomes of Task 1, one can train and/or test perception algorithms that deal with Moon images. For Task 2, one can produce as much data as desired to train and test AI algorithms that are anticipated to work in lunar conditions. All the outcomes of this work are available in a public repository for unfettered use and distribution.

Hyun Jin Park, Dhruuv Agarwal, Neng Chen et al.

This paper explores the use of TTS synthesized training data for KWS (keyword spotting) task while minimizing development cost and time. Keyword spotting models require a huge amount of training data to be accurate, and obtaining such training data can be costly. In the current state of the art, TTS models can generate large amounts of natural-sounding data, which can help reducing cost and time for KWS model development. Still, TTS generated data can be lacking diversity compared to real data. To pursue maximizing KWS model accuracy under the constraint of limited resources and current TTS capability, we explored various strategies to mix TTS data and real human speech data, with a focus on minimizing real data use and maximizing diversity of TTS output. Our experimental results indicate that relatively small amounts of real audio data with speaker diversity (100 speakers, 2k utterances) and large amounts of TTS synthesized data can achieve reasonably high accuracy (within 3x error rate of baseline), compared to the baseline (trained with 3.8M real positive utterances).

Bocheng Zou, Harry Zhang, Khailanii Slaton et al.

High-fidelity physics simulation is essential for closing the sim-to-real gap in robotics and complex mechanical systems. However, the computational overhead of high-fidelity engines often limits their use in data-intensive tasks like Reinforcement Learning (RL) and global optimization. We introduce Chrono-Gymnasium, a distributed computing framework that scales the high-fidelity multi-body dynamics of Project Chrono across large-scale computing clusters. Built upon the Ray framework, Chrono-Gymnasium provides a standardized Gymnasium interface, enabling seamless integration with modern machine learning libraries while providing built-in synchronization and messaging primitives for distributed execution. We demonstrate the framework's capabilities through two distinct case studies: (1) the training of an RL agent for autonomous robotic navigation in complex terrains, and (2) the Bayesian Optimization of a planetary lander's design parameters to ensure landing stability. Our results show that Chrono-Gymnasium reduces wall-clock time for high-fidelity simulations without sacrificing physical accuracy, offering a scalable path for the design and control of complex robotic systems.

Jingquan Wang, Andrew Negrut, Harry Zhang et al.

This contribution is concerned with the following issue: can pretrained large language models (LLMs) be refined and customized to the point where they become virtual assistants helping experts with the effective use of a simulation tool? In this case study, the ``simulation tool'' considered is PyChrono, an open source multi-physics dynamics engine for multibody systems. We present a framework for refining and customizing both open- and closed-source LLMs to harness the power of AI in generating scripts that perform PyChrono virtual experiments. We refine and customize several classes of LLMs through a process that leads to a quantifiable improvement in the quality of the generated PyChrono simulation scripts. These scripts can range from simple single-pendulum simulations to complex virtual experiments involving full vehicles on deformable terrain. While the generated scripts are rarely perfect, they often serve as strong starting points for the user to modify and improve on. Additionally, the LLM can answer specific API questions about the simulator, or recommend modeling approaches. The framework discussed is general and can be applied to lower the entry barrier for simulation tools associated with other application domains.

Jingquan Wang, Harry Zhang, Khailanii Slaton et al.

Recently, the integration of advanced simulation technologies with artificial intelligence (AI) is revolutionizing science and engineering research. ChronoLlama introduces a novel framework that customizes the open-source LLMs, specifically for code generation, paired with PyChrono for multi-physics simulations. This integration aims to automate and improve the creation of simulation scripts, thus enhancing model accuracy and efficiency. This combination harnesses the speed of AI-driven code generation with the reliability of physics-based simulations, providing a powerful tool for researchers and engineers. Empirical results indicate substantial enhancements in simulation setup speed, accuracy of the generated codes, and overall computational efficiency. ChronoLlama not only expedites the development and testing of multibody systems but also spearheads a scalable, AI-enhanced approach to managing intricate mechanical simulations. This pioneering integration of cutting-edge AI with traditional simulation platforms represents a significant leap forward in automating and optimizing design processes in engineering applications.

Hyun Jin Park, Dhruuv Agarwal, Neng Chen et al.

The keyword spotting (KWS) problem requires large amounts of real speech training data to achieve high accuracy across diverse populations. Utilizing large amounts of text-to-speech (TTS) synthesized data can reduce the cost and time associated with KWS development. However, TTS data may contain artifacts not present in real speech, which the KWS model can exploit (overfit), leading to degraded accuracy on real speech. To address this issue, we propose applying an adversarial training method to prevent the KWS model from learning TTS-specific features when trained on large amounts of TTS data. Experimental results demonstrate that KWS model accuracy on real speech data can be improved by up to 12% when adversarial loss is used in addition to the original KWS loss. Surprisingly, we also observed that the adversarial setup improves accuracy by up to 8%, even when trained solely on TTS and real negative speech data, without any real positive examples.

David Jin, Sushrut Karmalkar, Harry Zhang et al.

We investigate a variation of the 3D registration problem, named multi-model 3D registration. In the multi-model registration problem, we are given two point clouds picturing a set of objects at different poses (and possibly including points belonging to the background) and we want to simultaneously reconstruct how all objects moved between the two point clouds. This setup generalizes standard 3D registration where one wants to reconstruct a single pose, e.g., the motion of the sensor picturing a static scene. Moreover, it provides a mathematically grounded formulation for relevant robotics applications, e.g., where a depth sensor onboard a robot perceives a dynamic scene and has the goal of estimating its own motion (from the static portion of the scene) while simultaneously recovering the motion of all dynamic objects. We assume a correspondence-based setup where we have putative matches between the two point clouds and consider the practical case where these correspondences are plagued with outliers. We then propose a simple approach based on Expectation-Maximization (EM) and establish theoretical conditions under which the EM approach converges to the ground truth. We evaluate the approach in simulated and real datasets ranging from table-top scenes to self-driving scenarios and demonstrate its effectiveness when combined with state-of-the-art scene flow methods to establish dense correspondences.

Jingnan Shi, Rajat Talak, Harry Zhang et al.

We consider the problem of estimating object pose and shape from an RGB-D image. Our first contribution is to introduce CRISP, a category-agnostic object pose and shape estimation pipeline. The pipeline implements an encoder-decoder model for shape estimation. It uses FiLM-conditioning for implicit shape reconstruction and a DPT-based network for estimating pose-normalized points for pose estimation. As a second contribution, we propose an optimization-based pose and shape corrector that can correct estimation errors caused by a domain gap. Observing that the shape decoder is well behaved in the convex hull of known shapes, we approximate the shape decoder with an active shape model, and show that this reduces the shape correction problem to a constrained linear least squares problem, which can be solved efficiently by an interior point algorithm. Third, we introduce a self-training pipeline to perform self-supervised domain adaptation of CRISP. The self-training is based on a correct-and-certify approach, which leverages the corrector to generate pseudo-labels at test time, and uses them to self-train CRISP. We demonstrate CRISP (and the self-training) on YCBV, SPE3R, and NOCS datasets. CRISP shows high performance on all the datasets. Moreover, our self-training is capable of bridging a large domain gap. Finally, CRISP also shows an ability to generalize to unseen objects. Code and pre-trained models will be available on https://web.mit.edu/sparklab/research/crisp_object_pose_shape/.

Priya Sundaresan, Aditya Ganapathi, Harry Zhang et al.

We investigate the problem of pixelwise correspondence for deformable objects, namely cloth and rope, by comparing both classical and learning-based methods. We choose cloth and rope because they are traditionally some of the most difficult deformable objects to analytically model with their large configuration space, and they are meaningful in the context of robotic tasks like cloth folding, rope knot-tying, T-shirt folding, curtain closing, etc. The correspondence problem is heavily motivated in robotics, with wide-ranging applications including semantic grasping, object tracking, and manipulation policies built on top of correspondences. We present an exhaustive survey of existing classical methods for doing correspondence via feature-matching, including SIFT, SURF, and ORB, and two recently published learning-based methods including TimeCycle and Dense Object Nets. We make three main contributions: (1) a framework for simulating and rendering synthetic images of deformable objects, with qualitative results demonstrating transfer between our simulated and real domains (2) a new learning-based correspondence method extending Dense Object Nets, and (3) a standardized comparison across state-of-the-art correspondence methods. Our proposed method provides a flexible, general formulation for learning temporally and spatially continuous correspondences for nonrigid (and rigid) objects. We report root mean squared error statistics for all methods and find that Dense Object Nets outperforms baseline classical methods for correspondence, and our proposed extension of Dense Object Nets performs similarly.

Harry Zhang, Kurt Partridge, Pai Zhu et al.

Spoken Keyword Spotting (KWS) is the task of distinguishing between the presence and absence of a keyword in audio. The accuracy of a KWS model hinges on its ability to correctly classify examples close to the keyword and non-keyword boundary. These boundary examples are often scarce in training data, limiting model performance. In this paper, we propose a method to systematically generate adversarial examples close to the decision boundary by making insertion/deletion/substitution edits on the keyword's graphemes. We evaluate this technique on held-out data for a popular keyword and show that the technique improves AUC on a dataset of synthetic hard negatives by 61% while maintaining quality on positives and ambient negative audio data.

Harry Zhang, Luca Carlone

Understanding how humans interact with the surrounding environment, and specifically reasoning about object interactions and affordances, is a critical challenge in computer vision, robotics, and AI. Current approaches often depend on labor-intensive, hand-labeled datasets capturing real-world or simulated human-object interaction (HOI) tasks, which are costly and time-consuming to produce. Furthermore, most existing methods for 3D affordance understanding are limited to contact-based analysis, neglecting other essential aspects of human-object interactions, such as orientation (\eg, humans might have a preferential orientation with respect certain objects, such as a TV) and spatial occupancy (\eg, humans are more likely to occupy certain regions around an object, like the front of a microwave rather than its back). To address these limitations, we introduce \emph{H2OFlow}, a novel framework that comprehensively learns 3D HOI affordances -- encompassing contact, orientation, and spatial occupancy -- using only synthetic data generated from 3D generative models. H2OFlow employs a dense 3D-flow-based representation, learned through a dense diffusion process operating on point clouds. This learned flow enables the discovery of rich 3D affordances without the need for human annotations. Through extensive quantitative and qualitative evaluations, we demonstrate that H2OFlow generalizes effectively to real-world objects and surpasses prior methods that rely on manual annotations or mesh-based representations in modeling 3D affordance.

Harry Zhang, Luca Carlone

We introduce CUPS, a novel method for learning sequence-to-sequence 3D human shapes and poses from RGB videos with uncertainty quantification. To improve on top of prior work, we develop a method to generate and score multiple hypotheses during training, effectively integrating uncertainty quantification into the learning process. This process results in a deep uncertainty function that is trained end-to-end with the 3D pose estimator. Post-training, the learned deep uncertainty model is used as the conformity score, which can be used to calibrate a conformal predictor in order to assess the quality of the output prediction. Since the data in human pose-shape learning is not fully exchangeable, we also present two practical bounds for the coverage gap in conformal prediction, developing theoretical backing for the uncertainty bound of our model. Our results indicate that by taking advantage of deep uncertainty with conformal prediction, our method achieves state-of-the-art performance across various metrics and datasets while inheriting the probabilistic guarantees of conformal prediction.

David Jin, Harry Zhang, Kai Chang

We perform detailed theoretical analysis of an expectation-maximization-based algorithm recently proposed in for solving a variation of the 3D registration problem, named multi-model 3D registration. Despite having shown superior empirical results, did not theoretically justify the conditions under which the EM approach converges to the ground truth. In this project, we aim to close this gap by establishing such conditions. In particular, the analysis revolves around the usage of probabilistic tail bounds that are developed and applied in various instances throughout the course. The problem studied in this project stands as another example, different from those seen in the course, in which tail-bounds help advance our algorithmic understanding in a probabilistic way. We provide self-contained background materials on 3D Registration

Sitian Shen, Zilin Zhu, Linqian Fan et al.

Large pre-trained models have had a significant impact on computer vision by enabling multi-modal learning, where the CLIP model has achieved impressive results in image classification, object detection, and semantic segmentation. However, the model's performance on 3D point cloud processing tasks is limited due to the domain gap between depth maps from 3D projection and training images of CLIP. This paper proposes DiffCLIP, a new pre-training framework that incorporates stable diffusion with ControlNet to minimize the domain gap in the visual branch. Additionally, a style-prompt generation module is introduced for few-shot tasks in the textual branch. Extensive experiments on the ModelNet10, ModelNet40, and ScanObjectNN datasets show that DiffCLIP has strong abilities for 3D understanding. By using stable diffusion and style-prompt generation, DiffCLIP achieves an accuracy of 43.2\% for zero-shot classification on OBJ\_BG of ScanObjectNN, which is state-of-the-art performance, and an accuracy of 80.6\% for zero-shot classification on ModelNet10, which is comparable to state-of-the-art performance.

Dhruv Guliani, Lillian Zhou, Changwan Ryu et al.

Federated learning can be used to train machine learning models on the edge on local data that never leave devices, providing privacy by default. This presents a challenge pertaining to the communication and computation costs associated with clients' devices. These costs are strongly correlated with the size of the model being trained, and are significant for state-of-the-art automatic speech recognition models. We propose using federated dropout to reduce the size of client models while training a full-size model server-side. We provide empirical evidence of the effectiveness of federated dropout, and propose a novel approach to vary the dropout rate applied at each layer. Furthermore, we find that federated dropout enables a set of smaller sub-models within the larger model to independently have low word error rates, making it easier to dynamically adjust the size of the model deployed for inference.

Shivin Devgon, Jeffrey Ichnowski, Ashwin Balakrishna et al.

Orienting objects is a critical component in the automation of many packing and assembly tasks. We present an algorithm to orient novel objects given a depth image of the object in its current and desired orientation. We formulate a self-supervised objective for this problem and train a deep neural network to estimate the 3D rotation as parameterized by a quaternion, between these current and desired depth images. We then use the trained network in a proportional controller to re-orient objects based on the estimated rotation between the two depth images. Results suggest that in simulation we can rotate unseen objects with unknown geometries by up to 30° with a median angle error of 1.47° over 100 random initial/desired orientations each for 22 novel objects. Experiments on physical objects suggest that the controller can achieve a median angle error of 4.2° over 10 random initial/desired orientations each for 5 objects.

Harry Zhang, Jeffrey Ichnowski, Daniel Seita et al.

We explore how high-speed robot arm motions can dynamically manipulate cables to vault over obstacles, knock objects from pedestals, and weave between obstacles. In this paper, we propose a self-supervised learning framework that enables a UR5 robot to perform these three tasks. The framework finds a 3D apex point for the robot arm, which, together with a task-specific trajectory function, defines an arcing motion that dynamically manipulates the cable to perform tasks with varying obstacle and target locations. The trajectory function computes minimum-jerk motions that are constrained to remain within joint limits and to travel through the 3D apex point by repeatedly solving quadratic programs to find the shortest and fastest feasible motion. We experiment with 5 physical cables with different thickness and mass and compare performance against two baselines in which a human chooses the apex point. Results suggest that a baseline with a fixed apex across the three tasks achieves respective success rates of 51.7%, 36.7%, and 15.0%, and a baseline with human-specified, task-specific apex points achieves 66.7%, 56.7%, and 15.0% success rate respectively, while the robot using the learned apex point can achieve success rates of 81.7% in vaulting, 65.0% in knocking, and 60.0% in weaving. Code, data, and supplementary materials are available at https: //sites.google.com/berkeley.edu/dynrope/home.

Yahav Avigal, Samuel Paradis, Harry Zhang

Recent consumer demand for home robots has accelerated performance of robotic grasping. However, a key component of the perception pipeline, the depth camera, is still expensive and inaccessible to most consumers. In addition, grasp planning has significantly improved recently, by leveraging large datasets and cloud robotics, and by limiting the state and action space to top-down grasps with 4 degrees of freedom (DoF). By leveraging multi-view geometry of the object using inexpensive equipment such as off-the-shelf RGB cameras and state-of-the-art algorithms such as Learn Stereo Machine (LSM\cite{kar2017learning}), the robot is able to generate more robust grasps from different angles with 6-DoF. In this paper, we present a modification of LSM to graspable objects, evaluate the grasps, and develop a 6-DoF grasp planner based on Grasp-Quality CNN (GQ-CNN\cite{mahler2017dex}) that exploits multiple camera views to plan a robust grasp, even in the absence of a possible top-down grasp.

Khe Chai Sim, Françoise Beaufays, Arnaud Benard et al.

We study the effectiveness of several techniques to personalize end-to-end speech models and improve the recognition of proper names relevant to the user. These techniques differ in the amounts of user effort required to provide supervision, and are evaluated on how they impact speech recognition performance. We propose using keyword-dependent precision and recall metrics to measure vocabulary acquisition performance. We evaluate the algorithms on a dataset that we designed to contain names of persons that are difficult to recognize. Therefore, the baseline recall rate for proper names in this dataset is very low: 2.4%. A data synthesis approach we developed brings it to 48.6%, with no need for speech input from the user. With speech input, if the user corrects only the names, the name recall rate improves to 64.4%. If the user corrects all the recognition errors, we achieve the best recall of 73.5%. To eliminate the need to upload user data and store personalized models on a server, we focus on performing the entire personalization workflow on a mobile device.