$π$, But Make It Fly: Physics-Guided Transfer of VLA Models to Aerial Manipulation

Vision-Language-Action (VLA) models such as $π_0$ have demonstrated remarkable generalization across diverse fixed-base manipulators. However, transferring these foundation models to aerial platforms remains an open challenge due to the fundamental mismatch between the quasi-static dynamics of fixed-base arms and the underactuated, highly dynamic nature of flight. In this work, we introduce AirVLA, a system that investigates the transferability of manipulation-pretrained VLAs to aerial pick-and-place tasks. We find that while visual representations transfer effectively, the specific control dynamics required for flight do not. To bridge this "dynamics gap" without retraining the foundation model, we introduce a Payload-Aware Guidance mechanism that injects payload constraints directly into the policy's flow-matching sampling process. To overcome data scarcity, we further utilize a Gaussian Splatting pipeline to synthesize navigation training data. We evaluate our method through a cumulative 460 real-world experiments which demonstrate that this synthetic data is a key enabler of performance, unlocking 100% success in navigation tasks where directly fine-tuning on teleoperation data alone attains 81% success. Our inference-time intervention, Payload-Aware Guidance, increases real-world pick-and-place task success from 23% to 50%. Finally, we evaluate the model on a long-horizon compositional task, achieving a 62% overall success rate. These results suggest that pre-trained manipulation VLAs, with appropriate data augmentation and physics-informed guidance, can transfer to aerial manipulation and navigation, as well as the composition of these tasks.

Sparse Autoencoders Reveal Interpretable and Steerable Features in VLA Models

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

Observing and Controlling Features in Vision-Language-Action Models

Vision-Language-Action Models (VLAs) have shown remarkable progress towards embodied intelligence. While their architecture partially resembles that of Large Language Models (LLMs), VLAs exhibit higher complexity due to their multi-modal inputs/outputs and often hybrid nature of transformer and diffusion heads. This is part of the reason why insights from mechanistic interpretability in LLMs, which explain how the internal model representations relate to their output behavior, do not trivially transfer to VLA counterparts. In this work, we propose to close this gap by introducing and analyzing two main concepts: feature-observability and feature-controllability. In particular, we first study features that are linearly encoded in representation space, and show how they can be observed by means of a linear classifier. Then, we use a minimal linear intervention grounded in optimal control to accurately place internal representations and steer the VLA's output towards a desired region. Our results show that targeted, lightweight interventions can reliably steer a robot's behavior while preserving closed-loop capabilities. We demonstrate on different VLA architectures ($π_{0.5}$ and OpenVLA) through simulation experiments that VLAs possess interpretable internal structure amenable to online adaptation without fine-tuning, enabling real-time alignment with user preferences and task requirements.

Semantic-Metric Bayesian Risk Fields: Learning Robot Safety from Human Videos with a VLM Prior

Humans interpret safety not as a binary signal but as a continuous, context- and spatially-dependent notion of risk. While risk is subjective, humans form rational mental models that guide action selection in dynamic environments. This work proposes a framework for extracting implicit human risk models by introducing a novel, semantically-conditioned and spatially-varying parametrization of risk, supervised directly from safe human demonstration videos and VLM common sense. Notably, we define risk through a Bayesian formulation. The prior is furnished by a pretrained vision-language model. In order to encourage the risk estimate to be more human aligned, a likelihood function modulates the prior to produce a relative metric of risk. Specifically, the likelihood is a learned ViT that maps pretrained features, to pixel-aligned risk values. Our pipeline ingests RGB images and a query object string, producing pixel-dense risk images. These images that can then be used as value-predictors in robot planning tasks or be projected into 3D for use in conventional trajectory optimization to produce human-like motion. This learned mapping enables generalization to novel objects and contexts, and has the potential to scale to much larger training datasets. In particular, the Bayesian framework that is introduced enables fast adaptation of our model to additional observations or common sense rules. We demonstrate that our proposed framework produces contextual risk that aligns with human preferences. Additionally, we illustrate several downstream applications of the model; as a value learner for visuomotor planners or in conjunction with a classical trajectory optimization algorithm. Our results suggest that our framework is a significant step toward enabling autonomous systems to internalize human-like risk. Code and results can be found at https://riskbayesian.github.io/bayesian_risk/.

DexFruit: Dexterous Manipulation and Gaussian Splatting Inspection of Fruit

DexFruit is a robotic manipulation framework that enables gentle, autonomous handling of fragile fruit and precise evaluation of damage. Many fruits are fragile and prone to bruising, thus requiring humans to manually harvest them with care. In this work, we demonstrate by using optical tactile sensing, autonomous manipulation of fruit with minimal damage can be achieved. We show that our tactile informed diffusion policies outperform baselines in both reduced bruising and pick-and-place success rate across three fruits: strawberries, tomatoes, and blackberries. In addition, we introduce FruitSplat, a novel technique to represent and quantify visual damage in high-resolution 3D representation via 3D Gaussian Splatting (3DGS). Existing metrics for measuring damage lack quantitative rigor or require expensive equipment. With FruitSplat, we distill a 2D strawberry mask as well as a 2D bruise segmentation mask into the 3DGS representation. Furthermore, this representation is modular and general, compatible with any relevant 2D model. Overall, we demonstrate a 92% grasping policy success rate, up to a 20% reduction in visual bruising, and up to an 31% improvement in grasp success rate on challenging fruit compared to our baselines across our three tested fruits. We rigorously evaluate this result with over 630 trials. Please checkout our website at https://dex-fruit.github.io .

TensorTouch: Calibration of Tactile Sensors for High Resolution Stress Tensor and Deformation for Dexterous Manipulation

Advanced dexterous manipulation involving multiple simultaneous contacts across different surfaces, like pinching coins from ground or manipulating intertwined objects, remains challenging for robotic systems. Such tasks exceed the capabilities of vision and proprioception alone, requiring high-resolution tactile sensing with calibrated physical metrics. Raw optical tactile sensor images, while information-rich, lack interpretability and cross-sensor transferability, limiting their real-world utility. TensorTouch addresses this challenge by integrating finite element analysis with deep learning to extract comprehensive contact information from optical tactile sensors, including stress tensors, deformation fields, and force distributions at pixel-level resolution. The TensorTouch framework achieves sub-millimeter position accuracy and precise force estimation while supporting large sensor deformations crucial for manipulating soft objects. Experimental validation demonstrates 90% success in selectively grasping one of two strings based on detected motion, enabling new contact-rich manipulation capabilities previously inaccessible to robotic systems.

Next Best Sense: Guiding Vision and Touch with FisherRF for 3D Gaussian Splatting

We propose a framework for active next best view and touch selection for robotic manipulators using 3D Gaussian Splatting (3DGS). 3DGS is emerging as a useful explicit 3D scene representation for robotics, as it has the ability to represent scenes in a both photorealistic and geometrically accurate manner. However, in real-world, online robotic scenes where the number of views is limited given efficiency requirements, random view selection for 3DGS becomes impractical as views are often overlapping and redundant. We address this issue by proposing an end-to-end online training and active view selection pipeline, which enhances the performance of 3DGS in few-view robotics settings. We first elevate the performance of few-shot 3DGS with a novel semantic depth alignment method using Segment Anything Model 2 (SAM2) that we supplement with Pearson depth and surface normal loss to improve color and depth reconstruction of real-world scenes. We then extend FisherRF, a next-best-view selection method for 3DGS, to select views and touch poses based on depth uncertainty. We perform online view selection on a real robot system during live 3DGS training. We motivate our improvements to few-shot GS scenes, and extend depth-based FisherRF to them, where we demonstrate both qualitative and quantitative improvements on challenging robot scenes. For more information, please see our project page at https://armlabstanford.github.io/next-best-sense.

SAFER-Splat: A Control Barrier Function for Safe Navigation with Online Gaussian Splatting Maps

SAFER-Splat (Simultaneous Action Filtering and Environment Reconstruction) is a real-time, scalable, and minimally invasive action filter, based on control barrier functions, for safe robotic navigation in a detailed map constructed at runtime using Gaussian Splatting (GSplat). We propose a novel Control Barrier Function (CBF) that not only induces safety with respect to all Gaussian primitives in the scene, but when synthesized into a controller, is capable of processing hundreds of thousands of Gaussians while maintaining a minimal memory footprint and operating at 15 Hz during online Splat training. Of the total compute time, a small fraction of it consumes GPU resources, enabling uninterrupted training. The safety layer is minimally invasive, correcting robot actions only when they are unsafe. To showcase the safety filter, we also introduce SplatBridge, an open-source software package built with ROS for real-time GSplat mapping for robots. We demonstrate the safety and robustness of our pipeline first in simulation, where our method is 20-50x faster, safer, and less conservative than competing methods based on neural radiance fields. Further, we demonstrate simultaneous GSplat mapping and safety filtering on a drone hardware platform using only on-board perception. We verify that under teleoperation a human pilot cannot invoke a collision. Our videos and codebase can be found at https://chengine.github.io/safer-splat.

Splat-MOVER: Multi-Stage, Open-Vocabulary Robotic Manipulation via Editable Gaussian Splatting

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 latent codes for language semantics and grasp affordance into the 3D scene. ASK-Splat enables geometric, semantic, and affordance understanding of 3D scenes, which is critical for 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 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, as well as in four multi-stage manipulation tasks using the edited scene to reflect scene changes due to prior manipulation stages, which is not possible with the existing baselines. Code for this project and a link to the project page will be made available soon.

$\textbf{Splat-MOVER}$: Multi-Stage, Open-Vocabulary Robotic Manipulation via Editable Gaussian Splatting

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) $\textit{ASK-Splat}$, a GSplat representation that distills latent codes for language semantics and grasp affordance into the 3D scene. ASK-Splat enables geometric, semantic, and affordance understanding of 3D scenes, which is critical for many robotics tasks; (ii) $\textit{SEE-Splat}$, 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) $\textit{Grasp-Splat}$, a grasp generation module that uses ASK-Splat and SEE-Splat to propose 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, as well as in four multi-stage manipulation tasks using the edited scene to reflect scene changes due to prior manipulation stages, which is not possible with the existing baselines. Code for this project and a link to the project page will be made available soon.