Monroe Kennedy

Aiden Swann, Lachlain McGranahan, Hugo Buurmeijer et al.

Vision-Language-Action (VLA) models have emerged as a promising approach for general-purpose robot manipulation. However, their generalization is inconsistent: while these models can perform impressively in some settings, fine-tuned variants often fail on novel objects, scenes, and instructions. We apply mechanistic interpretability techniques to better understand the inner workings of VLA models. To probe internal representations, we train Sparse Autoencoders (SAEs) on hidden layer activations of the VLA. SAEs learn a sparse dictionary whose features act as a compact, interpretable basis for the model's computation. We find that the large majority of extracted SAE features correspond to memorized sequences from specific training demonstrations. However, some features correspond to interpretable, general, and steerable motion primitives and semantic properties, offering a promising glimpse toward VLA generalizability. We propose a metric to categorize features according to whether they represent generalizable transferable primitives or episode-specific memorization. We validate these findings through steering experiments on the LIBERO benchmark. We show that individual SAE features causally influence robot behavior. Steering general features induces behaviors consistent with their semantic meaning and can be applied across tasks and scenes. This work provides the first mechanistic evidence that VLAs can learn generalizable features across tasks and scenes. We observe that supervised fine-tuning on small robotics datasets disproportionately amplifies memorization. In contrast, training on larger, more diverse datasets (e.g., DROID) or using knowledge insulation promotes more general features. We provide an open-source codebase and user-friendly interface for activation collection, SAE training, and feature steering. Our project page is located at http://drvla.github.io

Eley Ng, Ziang Liu, Monroe Kennedy

Cooperative table-carrying is a complex task due to the continuous nature of the action and state-spaces, multimodality of strategies, and the need for instantaneous adaptation to other agents. In this work, we present a method for predicting realistic motion plans for cooperative human-robot teams on the task. Using a Variational Recurrent Neural Network (VRNN) to model the variation in the trajectory of a human-robot team across time, we are able to capture the distribution over the team's future states while leveraging information from interaction history. The key to our approach is leveraging human demonstration data to generate trajectories that synergize well with humans during test time in a receding horizon fashion. Comparison between a baseline, sampling-based planner RRT (Rapidly-exploring Random Trees) and the VRNN planner in centralized planning shows that the VRNN generates motion more similar to the distribution of human-human demonstrations than the RRT. Results in a human-in-the-loop user study show that the VRNN planner outperforms decentralized RRT on task-related metrics, and is significantly more likely to be perceived as human than the RRT planner. Finally, we demonstrate the VRNN planner on a real robot paired with a human teleoperating another robot.

Weizhuo Wang, Yanjie Ze, C. Karen Liu et al.

We present EgoNav, a system that enables a humanoid robot to traverse diverse, unseen environments by learning entirely from 5 hours of human walking data, with no robot data or finetuning. A diffusion model predicts distributions of plausible future trajectories conditioned on past trajectory, a 360 deg visual memory fusing color, depth, and semantics, and video features from a frozen DINOv3 backbone that capture appearance cues invisible to depth sensors. A hybrid sampling scheme achieves real-time inference in 10 denoising steps, and a receding-horizon controller selects paths from the predicted distribution. We validate EgoNav through offline evaluations, where it outperforms baselines in collision avoidance and multi-modal coverage, and through zero-shot deployment on a Unitree G1 humanoid across unseen indoor and outdoor environments. Behaviors such as waiting for doors to open, navigating around crowds, and avoiding glass walls emerge naturally from the learned prior. We will release the dataset and trained models. Our website: https://egonav.weizhuowang.com

Shivani Guptasarma, Matthew Strong, Honghao Zhen et al.

J-PARSE is an algorithm for smooth first-order inverse kinematic control of a serial manipulator near kinematic singularities. The commanded end-effector velocity is interpreted component-wise, according to the available mobility in each dimension of the task space. First, a substitute "Safety" Jacobian matrix is created, keeping the aspect ratio of the manipulability ellipsoid above a threshold value. The desired motion is then projected onto non-singular and singular directions, and the latter projection scaled down by a factor informed by the threshold value. A right-inverse of the non-singular Safety Jacobian is applied to the modified command. In the absence of joint limits and collisions, this ensures safe transition into and out of low-rank configurations, guaranteeing asymptotic stability for reaching target poses within the workspace, and stability for those outside. Velocity control with J-PARSE is benchmarked against approaches from the literature, and shows high accuracy in reaching and leaving singular target poses. By expanding the available workspace of manipulators, the algorithm finds applications in teleoperation, servoing, and learning. Videos and code are available at https://jparse-manip.github.io/.

Ola Shorinwa, Johnathan Tucker, Aliyah Smith et al.

We present Splat-MOVER, a modular robotics stack for open-vocabulary robotic manipulation, which leverages the editability of Gaussian Splatting (GSplat) scene representations to enable multi-stage manipulation tasks. Splat-MOVER consists of: (i) ASK-Splat, a GSplat representation that distills semantic and grasp affordance features into the 3D scene. ASK-Splat enables geometric, semantic, and affordance understanding of 3D scenes, which is critical in many robotics tasks; (ii) SEE-Splat, a real-time scene-editing module using 3D semantic masking and infilling to visualize the motions of objects that result from robot interactions in the real-world. SEE-Splat creates a "digital twin" of the evolving environment throughout the manipulation task; and (iii) Grasp-Splat, a grasp generation module that uses ASK-Splat and SEE-Splat to propose affordance-aligned candidate grasps for open-world objects. ASK-Splat is trained in real-time from RGB images in a brief scanning phase prior to operation, while SEE-Splat and Grasp-Splat run in real-time during operation. We demonstrate the superior performance of Splat-MOVER in hardware experiments on a Kinova robot compared to two recent baselines in four single-stage, open-vocabulary manipulation tasks and in four multi-stage manipulation tasks, using the edited scene to reflect changes due to prior manipulation stages, which is not possible with existing baselines. Video demonstrations and the code for the project are available at https://splatmover.github.io.

Aliyah Smith, Monroe Kennedy

Robot sound, encompassing both consequential operational noise and intentionally designed auditory cues, plays an important role in human-robot interaction (HRI). Developing a deeper understanding of how robot sounds influence human experience, and how technologies such as augmented reality (AR) modulate these effects, can enable the design of more socially acceptable robots and more effective, intuitive human-robot interfaces. In this work, we present a three-part mixed-methods study (N = 51) that investigates (i) the effects of consequential robot sounds on human perception under varying degrees of physical colocation, (ii) human accuracy in localizing spatial audio cues delivered via augmented reality, and (iii) the use of augmented spatial audio cues for functional and transformative communication during collaborative handover tasks, in comparison to non-AR sound designs. Contrary to prior findings, our results indicate that the consequential sounds of a Kinova Gen3 manipulator did not negatively affect participants' perceptions of the robot. Participants demonstrated high accuracy in localizing lateral spatial cues, whereas frontal cues proved more challenging, delineating conditions under which spatial auditory feedback is most effective. Qualitative findings further reveal that augmented spatial audio cues can simultaneously convey task-relevant information while fostering a sense of warmth and reducing user discomfort during interaction. Together, these findings elucidate the perceptual effects of consequential robot sound and position sound, particularly augmented spatial audio, as a meaningful yet underutilized design resource for human-robot interaction.

Yang You, Won Kyung Do, Aiden Swann et al.

Simulating optical tactile sensors presents significant challenges due to their high deformability and intricate optical properties. To address these issues and enable a physically accurate simulation, we propose DOT-Sim: Differentiable Optical Tactile Simulation. Unlike prior simulators that rely on simplified models of deformable sensors, DOT-Sim accurately captures the physical behavior of soft sensors by modeling them as elastic materials using the Material Point Method (MPM). DOT-Sim enables rapid calibration of optical tactile sensor simulation using a small number of demonstrations within minutes, which is substantially faster than existing methods. Compared to current baselines, our approach supports much larger and non-linear deformations. To handle the optical aspect, we propose a novel approach to simulating optical responses by learning a residual image relative to the real-world idle state. We validate the physical and visual realism of our method through a series of zero-shot sim-to-real tasks. Our experiments show that DOT-Sim (1) accurately replicates the physical dynamics of a DenseTact optical tactile sensor in reality, (2) generates realistic optical outputs in contact-rich scenarios, (3) enables direct deployment of simulation-trained classifiers in the real world, achieving 85% classification accuracy on challenging objects and 90% accuracy in embedded tumor-type detection, and (4) allows precise trajectory following with a policy trained from demonstrations in simulation, with an average error of less than 0.9 mm.

Weizhuo Wang, C. Karen Liu, Monroe Kennedy

Wearable collaborative robots stand to assist human wearers who need fall prevention assistance or wear exoskeletons. Such a robot needs to be able to constantly adapt to the surrounding scene based on egocentric vision, and predict the ego motion of the wearer. In this work, we leveraged body-mounted cameras and sensors to anticipate the trajectory of human wearers through complex surroundings. To facilitate research in ego-motion prediction, we have collected a comprehensive walking scene navigation dataset centered on the user's perspective. We then present a method to predict human motion conditioning on the surrounding static scene. Our method leverages a diffusion model to produce a distribution of potential future trajectories, taking into account the user's observation of the environment. To that end, we introduce a compact representation to encode the user's visual memory of the surroundings, as well as an efficient sample-generating technique to speed up real-time inference of a diffusion model. We ablate our model and compare it to baselines, and results show that our model outperforms existing methods on key metrics of collision avoidance and trajectory mode coverage.

Aiden Swann, Matthew Strong, Won Kyung Do et al.

In this work, we propose a novel method to supervise 3D Gaussian Splatting (3DGS) scenes using optical tactile sensors. Optical tactile sensors have become widespread in their use in robotics for manipulation and object representation; however, raw optical tactile sensor data is unsuitable to directly supervise a 3DGS scene. Our representation leverages a Gaussian Process Implicit Surface to implicitly represent the object, combining many touches into a unified representation with uncertainty. We merge this model with a monocular depth estimation network, which is aligned in a two stage process, coarsely aligning with a depth camera and then finely adjusting to match our touch data. For every training image, our method produces a corresponding fused depth and uncertainty map. Utilizing this additional information, we propose a new loss function, variance weighted depth supervised loss, for training the 3DGS scene model. We leverage the DenseTact optical tactile sensor and RealSense RGB-D camera to show that combining touch and vision in this manner leads to quantitatively and qualitatively better results than vision or touch alone in a few-view scene syntheses on opaque as well as on reflective and transparent objects. Please see our project page at http://armlabstanford.github.io/touch-gs

Won Kyung Do, Monroe Kennedy

Increasing the performance of tactile sensing in robots enables versatile, in-hand manipulation. Vision-based tactile sensors have been widely used as rich tactile feedback has been shown to be correlated with increased performance in manipulation tasks. Existing tactile sensor solutions with high resolution have limitations that include low accuracy, expensive components, or lack of scalability. In this paper, an inexpensive, scalable, and compact tactile sensor with high-resolution surface deformation modeling for surface reconstruction of the 3D sensor surface is proposed. By measuring the image from the fisheye camera, it is shown that the sensor can successfully estimate the surface deformation in real-time (1.8ms) by using deep convolutional neural networks. This sensor in its design and sensing abilities represents a significant step toward better object in-hand localization, classification, and surface estimation all enabled by high-resolution shape reconstruction.