Researchers at EPFL's MICROBS Lab have developed sound-powered machines by utilizing the Helmholtz resonance phenomenon in hollow structures. These acoustic resonators convert sound waves into directional thrust and controlled motion, acting as a propulsion system. The team created miniature boats and microfliers that use tuned cavities to generate movement and lift. These devices can be fabricated from various materials, enabling extremely small and lightweight designs. This work paves the way for future sound-responsive robotics capable of changing shape in response to sound.
Neura Robotics acquired Adlatus Robotics, an established provider of autonomous cleaning and sweeping robots, to expand its physical AI ecosystem. The acquisition aims to equip Adlatus systems with advanced sensor and AI capabilities to evolve them into intelligent agents within the Neuraverse. This integration combines Adlatus's real-world operational knowledge with Neura's physical AI infrastructure. The goal is to develop physical AI by allowing robots to understand and adapt to their environments rather than just following predetermined routes. Ultimately, the partnership seeks to build a new European industry by merging industrial expertise with artificial intelligence.
Consumer robotics will evolve from a vision of a single humanoid to a collection of specialized, beautifully designed objects, much like the "Beauty and the Beast" theme. Success will depend less on robotic capability and more on creating products that embody high-frequency daily rituals. Consumers are motivated by owning these rituals—such as skincare or cleaning—rather than simply automating chores. Winning companies will focus on aesthetics and emotional connection, making robots look like furniture rather than industrial machines. This shift is driven by converging advancements in AI, cost reduction, and consumer acceptance of intelligent hardware in the home.
Musical humanoid robots are proposed as novel experimental interfaces for investigating music-evoked emotions. Current research is limited because it relies on pre-recorded stimuli, failing to capture the multimodal and interactive nature of real musical experience. Humanoid robots are suitable because they allow for the parametric control of acoustic and non-acoustic variables. A case study on the WAS-5 robot demonstrated reproducible control over these variables necessary for music-emotion experiments. Consequently, these robots offer a methodological platform for future controlled investigations of music-evoked emotions.
Bidirectional endoscopic control involves an ambiguity in actions when the requested phase changes before the visual scene does, which the authors term intent aliasing. EndoLIFT is a vision-language-action policy that conditions continuous action generation using a latent-conditioned rectified flow. The model uses RGB input, a language instruction, and the previous state, along with a 32-D trajectory latent. EndoLIFT significantly improves navigation accuracy and reduces wrong-direction advances compared to models without latent conditioning. Evaluation shows language selection separates intent from trajectory latent contributions to directional correctness.
A co-design framework is proposed to simultaneously develop data collection and robotic execution devices by incorporating platform constraints into the design. This leads to the Koala Gripper system, which is a data capture device and robotic gripper platform. The design features a novel force-optimized linkage mechanism and user-centered ergonomics, resulting in improved dexterity compared to parallel jaw grippers. The actuated fingers are backdrivable with low mass, capable of secure grasps and precise manipulation. The platform is validated through an end-to-end data collection and policy execution pipeline using learning from demonstration.
Logic-VLA is a Vision-Language-Action model that conditions its behavior on Signal Temporal Logic (STL) specifications provided at inference time. It employs an STL encoder to capture temporal logic semantics for policy adaptation. The adaptation involves supervised fine-tuning followed by preference optimization using a flow-matching surrogate. This method improves formal requirement satisfaction by up to 40.7 percentage points over a base policy. Logic-VLA demonstrates the ability to adapt behavior to varying formal requirements without needing separate policies for each specification.
The paper proposes a unified framework combining differential flatness-based trajectory generation with Nonlinear Model Predictive Control (NMPC) for fixed-wing aerial systems. This approach integrates computationally efficient planning with predictive, constraint-aware control. A wind-aware sampling strategy is introduced within the NMPC to proactively account for wind disturbances and enforce aerodynamic constraints. The framework was validated through simulations and real-world experiments, showing improved tracking accuracy and robustness. Specifically, the wind-aware strategy enhances performance under strong wind conditions.
Imitation learning in robotics relies on human demonstrations, but removing them efficiently is challenging. Existing unlearning metrics do not account for what is removed from a policy during closed-loop action. The authors propose a retrain-calibrated audit assessing demonstration unlearning along two axes: behavior divergence and evidence detectability. This audit combines these axes to hypothesize joint retraining consistency against multiple independent retrains. Experimental results show that edits can affect behavior and evidence in different ways, demonstrating the value of this new framework.
The research synthesizes natural sit-to-stand (STS) motion for humanoids using reinforcement learning without prior demonstrations. A Proximal Policy Optimisation policy is trained using three components: a force/chair-height curriculum, random sampling of initial/target poses, and biomechanically inspired rewards. This guided assistance ensures the policy masters trajectories at each chair height before progressing to more difficult ones, preventing generalization collapse. The policy achieved over 97% balanced-standing success across eight chair heights on a deterministic evaluator. This method enables the robot to execute smooth, generalized STS motions from deep postures.
A reinforcement learning framework is presented to generate stand-up motion for a humanoid robot on deformable soft ground using human demonstration from hard ground. Terrain compliance is modeled using MuJoCo's soft-contact parameters. The reward system includes tracking reference motion and achieving explicit recovery objectives like pelvis height and torso uprightness. Training adapts the policy to delayed support force generation during contact, successfully completing the stand-up task on both hard and soft surfaces. Explicit recovery rewards were shown to be essential for successful stand-up, as reference tracking alone was insufficient.
A vision-only state estimation system is presented for X-configuration quadcopters without Inertial Measurement Units. It uses synchronized stereo images and motor commands to estimate the body-frame pose, velocity, and other states via an extended Kalman filter. The system utilizes stationary scene points as implicit inertial references and performs feature tracking and matching across cameras. A sparse 3D point cloud with per-point estimates is also generated through a full bundle adjustment. The output provides a body-frame state estimate and sparse scene flow, independent of GPS or IMU.
Virtual scenario-based testing is crucial for validating automated driving systems, but executing large test suites is difficult to scale. A framework was developed to automate the build, deployment, and distributed execution of CARLA-based scenario tests on a Kubernetes cluster. This involves packaging ROS 2 applications as Helm charts and using Argo Workflows to manage the distributed execution of OpenSCENARIO test cases. Evaluation showed that this distributed workflow significantly reduces end-to-end execution time, achieving up to an eightfold speedup compared to sequential testing. The method provides a foundation for scalable simulation testing and generating traceable evidence for safety arguments.