Manipulation, locomotion, sim-to-real transfer and autonomous driving: learning systems that have to survive physics.
arXiv:2609.18125v1 Announce Type: new Abstract: Humans often observe others before interacting and adjust their behavior accordingly. Robot navigation in crowds, however, often represents pedestrians...
The paper surveys Generative Physical Artificial Intelligence (GPAI), a field where large foundation models are integrated with physical robots. It introduces a taxonomy of five approaches—Robot Foundation Models, Vision‑Language Action models, Large Behavior Models, Diffusion Policy Models, and World Foundation Models—and discusses how they complement each other across domains such as autonomous vehicles, industrial automation, healthcare robotics, and humanoid systems. The review highlights performance gains, data‑efficient learning, sim‑to‑real transfer, edge‑compatible architectures, and safety frameworks as key research directions.
Mem2Ego introduces a vision‑language model for embodied navigation that combines global memory with egocentric visual inputs. By adaptively retrieving task‑relevant cues from a global memory module and aligning them with local perception, the framework improves spatial reasoning and decision‑making over long horizons. The method outperforms prior state‑of‑the‑art approaches on the HSSD and HM3D benchmarks and shows strong performance on a real robot.
arXiv:2609.17628v1 Announce Type: cross Abstract: Active perception allows autonomous agents to select their viewpoints rather than passively process the viewpoints given to them, enabling them to ta...
The paper introduces MVLGeo, a unified framework for cross-view object geo-localization that combines multiple viewpoints into a single model. It employs Vision‑Language Reranking to use contextual text from the query view, a multi‑view Mixture‑of‑Experts architecture to share knowledge and reduce redundancy, and an adaptive elliptical prior for positional encoding. Experiments on CVOGL benchmarks show that MVLGeo achieves state‑of‑the‑art performance and robustness to input degradation.
The paper presents a compact multimodal policy trained via behavioral cloning to drive autonomously in the CARLA simulator. Using five‑frame histories of RGB images, LiDAR, telemetry, and lane waypoints, the 1.36‑million‑parameter model predicts throttle, brake, and steering at 20 Hz. Trained on 236,882 windows (≈3.3 hours of driving) from 448 captures, the policy drives for hours on both training and unseen routes without collisions, demonstrating qualitative transfer and recovery from large trajectory deviations.
The paper introduces Market Signal Injection (MSI), an attack that alters how market data is formatted or described—without changing its numerical values—to influence large language model (LLM) pricing agents. Experiments on nine open‑weight and three proprietary models in simulated duopoly and triopoly markets show that sentiment‑based formatting changes cause significant shifts in firm behavior, profits, and consumer surplus. The study also demonstrates that model susceptibility varies across families, that larger models are not always more robust, and that techniques such as input canonicalization and decision boundary anchoring can partially mitigate these attacks.
The paper re‑examines the impact of removing the encoder from Action Chunking Transformers (ACT), a model used for robot manipulation learning. Contrary to the original claim that encoder removal drops success rates from 35% to 2%, the authors find no such dramatic decline in their re‑runs, though minor variations remain uncertain. They attribute the discrepancy to factors like training length and checkpoint selection, and note that the encoder’s latent variable offers little reconstruction benefit on the tested benchmark, while its removal speeds up training.
The paper investigates how to balance model size and fine‑tuning strategy for UAV audio classification. Using a dataset of 3,100 clips across 31 drone classes, it compares transformer and convolutional backbones under full fine‑tuning, classifier‑only fine‑tuning, and four parameter‑efficient fine‑tuning methods. Results show that selective batch‑norm tuning of EfficientNet‑B7 yields the best accuracy (97.65%) while updating less than 0.5% of parameters, and that lightweight CNNs generally outperform transformers in both accuracy and efficiency.
The paper introduces LDE, a framework that uses collaborative perception (CP) to generate high‑quality pseudo‑labels for unsupervised model adaptation in autonomous driving. It tackles communication limits, field‑of‑view mismatches, and label unreliability through selective feature sharing, FoV filtering, and curriculum learning. Experiments on 3D object detection show LDE surpasses pre‑trained models and existing adaptation methods.
LiteViLNet is a lightweight RGB‑geometry fusion network for road segmentation that uses a MobileNetV3 RGB encoder and a tiny depth‑wise‑separable geometry encoder. Its multi‑scale fusion module enhances modality‑specific features, performs cross‑modal interaction, and applies adaptive gating, while a depth‑wise large‑kernel bridge expands contextual support with minimal overhead. The U‑Net‑style decoder is trained with deep supervision, achieving state‑of‑the‑art performance on KITTI and ORFD benchmarks and running at up to 68.73 FPS on a Jetson Orin NX with TensorRT FP16.
The paper introduces HAP, a Hand-Driven Active Perception framework that predicts future six‑degree‑of‑freedom head motion in egocentric settings by conditioning on observed hand motion and inferred target context. HAP constructs a Predictive Target‑Centric Amodal Occlusion Graph to model current and potential occlusions among candidate objects, fuses this with hand and head motion history, and blends the learned trajectory with a constant‑velocity prior. Experiments on a public dataset and a newly released Bottle RGB‑D dataset demonstrate that HAP outperforms baseline methods in head‑motion prediction, highlighting the importance of hand‑driven intention and dynamic occlusion reasoning.
PhysVGGT is a feed‑forward model that predicts dense maps of friction coefficient, Shore hardness, Young's modulus, and density, along with object‑level mass, from a single RGB image in one forward pass. It treats physical property estimation as a dense per‑pixel prediction problem, using a visual geometry transformer to extract geometry‑aware tokens and separate dense and global prediction branches. A scalable pseudo‑label generation pipeline enables large‑scale weakly supervised training, and the model achieves state‑of‑the‑art performance on the ABO‑500 dataset while running 27× faster than previous methods.
AeroWeaver is a new embodied‑agent harness that integrates large language model (LLM) decision making with the executable skills of individual UAVs, enabling distributed, adaptive swarm execution. It connects semantic mission decisions to governed skills, organizes role‑conditioned local agents for coordination, and refines skill selection online using role‑indexed state‑action‑reward experience. Experiments demonstrate that AeroWeaver maintains valid skill execution without a central joint‑action generator and supports reward‑guided, training‑free adaptive learning from accumulated execution experience.
The paper introduces rMuscle, a real‑time Vision‑Language‑Action inference framework that mimics human muscle memory to accelerate robotic decision making. By exploiting repeated task similarity, rMuscle uses a dual‑phase cache: a Context Cache reuses visual‑token outputs and an Action Cache reuses neuron activation patterns, reducing computation and weight accesses. Experiments on RTX 4090 and Jetson Thor show 1.29–1.42× speedups on LIBERO, RoboTwin, and physical manipulation tasks while preserving success rates on real robots.
arXiv:2608.13014v2 Announce Type: replace Abstract: Understanding hand-object interaction from egocentric vision is essential for modeling how people physically engage with the surrounding world. Yet...
SynAgent is a framework that uses large language model agents to run autonomous experiments while building an explicit, revisable understanding of the synthesis process. Unlike traditional black‑box optimizers, SynAgent generates analysis skills on the fly and reasons multimodally over data such as X‑ray diffraction patterns and electron micrographs. In an 18‑experiment campaign on LiCoO₂ thin‑film deposition, SynAgent produced highly crystalline films and uncovered a sharp temperature threshold and optimal growth window (650–690 °C) for crystallization.
PACT‑WAM is a world‑action model that simultaneously predicts a 16‑step action trajectory and its corresponding visual forecast for robot manipulation. It uses a hierarchical history encoder that compresses past observations into fewer tokens, reducing processing cost by 75% compared to dense encoding. The model’s shared flow module updates action and visual states jointly, and a TiTok‑VAE decoder reconstructs multi‑view future images, which are then used by a vision‑language component (Proposal Review) to improve execution‑prefix selection and proposal rejection, boosting success rates on several benchmarks.
Changepoint-Aware World Models (CAWM) is a DreamerV3 agent that detects abrupt dynamics shifts in a robot’s environment using an online CUSUM test on internal prediction error. Upon detection, CAWM selectively forgets stale replay data while preserving the learned representation, enabling rapid recovery from shifts such as doubled gravity or halved actuator gain. Experiments on simulated locomotion show CAWM recovers faster than passive retraining and outperforms a baseline that respawns a fresh dynamics model, achieving significant return gains in the first 30k post‑shift frames.
This paper introduces an energy-aware approach to robotic manipulation by defining a joint-space mechanical-work proxy based on joint torque and angular displacement. A differentiable energy predictor is trained to estimate this work from robot states and actions, enabling it to serve as a regularizer that fine‑tunes a pretrained manipulation policy. Applied to RVT‑2 on RLBench, the method reduces average mechanical work from 208.8 J to 204.4 J (a 2.1 % drop) while slightly improving task success from 86.2 % to 86.9 % across 12 manipulation tasks.