Blattner awarded Built Robotics a $75 million contract. This agreement is specifically for the development of physical AI applications. These autonomous systems are intended to assist in solar construction efforts. The goal of this expansion is to help address growing energy demands. Information regarding this partnership was initially published on The Robot Report.
Ruggedized design has transitioned from a luxury to an absolute necessity in robotics. As autonomous systems venture outside of controlled settings, operational reliability demands robust construction. Incorporating rugged features is now a prerequisite for ensuring functional robot performance. This growing need underscores the critical importance of durability in modern mobile robotics.
Humanoid robots dominated the focus in June 2026. Developments included significant corporate milestones and new operational deployments. Several companies reached key production achievements with their humanoid models. Coverage detailing these top advancements was published by The Robot Report. These updates marked a high level of activity within the field during that month.
Apptronik has unveiled Apollo 2, which functions as a comprehensive data collection and training platform. This system is designed to enable continuous learning through real-world deployment. Alongside the Apollo 2 unveiling, Apptronik also revealed its flagship data collection and training facility. Details of this announcement were initially published on The Robot Report.
Developing advanced technology often leaves wiring integration as an afterthought, creating late-stage design conflicts despite perfect mechanical or software execution. Pre-existing, off-the-shelf components frequently fail to accommodate the specific spatial and functional demands of new systems. Custom harness solutions are necessary because they allow for precise integration within limited physical footprints. These custom designs resolve complex fitment issues that standardized wiring cannot manage effectively. Consequently, bespoke harnesses become essential for realizing next-generation technological advancements.
Modern robotics increasingly requires precise and energy-efficient motion control across various industries. Brushless DC (BLDC) motors are fundamental components due to their inherent high efficiency and reliability. Motor controllers are critical for determining the actual performance achieved by these motors. The interaction between motor type, controller design, and operational parameters significantly dictates system accuracy. Optimal motor control is therefore essential for maximizing robotic system functionality.
EmbodimentSemantic introduces a novel dataset and benchmark to address the spatial grounding limitations in vision-language-action (VLA) systems for robotics. It models scenes using directed object-relation-object triplets, explicitly capturing spatial relationships like support or containment. The dataset utilizes real-world manipulation observations from the SO101 robot arm, complemented by a simulator-grounded LIBERO benchmark with extensive view variations and automatically derived ground-truth relations. Testing reveals that existing Vision-Language Models often struggle with accurately predicting depth-aware and viewpoint-dependent spatial structures. This framework offers a unified method for diagnosing spatial grounding deficits in VLM perception for embodied manipulation tasks.
Service robots' object search efficiency depends on spatial priors, which are challenged by variable resident placement behaviors. Proposing PerSim, the authors developed a rigidity-gated hybrid policy that integrates trait-conditioned priors only when object placement is highly variable. To achieve scalability, they used a human-calibrated simulation pipeline to generate and validate transitions using continuous Big Five personality vectors as input. In user studies (N=200), personalization proved beneficial primarily for low-rigidity objects compared to the population-frequency baseline. Testing confirmed PerSim's efficacy by showing measurable search cost reductions in a simulated home digital twin environment.
Introduces GPAC, a four-layer hierarchical architecture for decentralized aerial transport using quadrotors to carry a cable-suspended payload. It achieves coordination through implicit mechanisms where each drone estimates its load share locally, eliminating the need for central communication or prior knowledge of system parameters like $N$ or total mass. The framework reconstructs payload position from individual agent measurements and uses only neighbor-position broadcasting for collision avoidance. GPAC integrates geometric control on the full nonlinear manifold, anti-swing regulation, wind rejection via an observer, and concurrent mass estimation. Safety is ensured by a priority-ordered Control Barrier Function (CBF)-inspired filter that maintains operational risk margins even during single constraint activation.
Develops FLYNN, a recurrent neural network structured based on the fruit fly brain connectome. Demonstrates that this architecture can perform vision-based navigation in MuJoCo, matching established networks' performance. A key advantage is its superior resistance to out-of-distribution data and sensory loss without retraining. Specifically, FLYNN remained functional with total vision loss while comparable hand-crafted models failed. The high representational modularity of FLYNN suggests a novel pathway for creating resilient artificial agents modeled after biology.
An invariant extended Kalman filter (IEKF) is developed for state estimation of serial rigid manipulators using the SE(3) Lie group formulation. This method leverages the group-affine property, allowing the Riccati equation to govern the true error covariance. It employs a physically separated noise model where accelerometer data provides translational twist and gyroscope noise contributes a state-dependent Coriolis term. The filter is structured as a modular chain of per-link IEKFs with linear computational cost relative to link count. Stability is proven via an exponential ultimate boundedness using a Lie algebra Lyapunov function, yielding a scalable stability certificate.
Researchers analyzed 1,219 sustained deceleration events from Argoverse 2 data to categorize urban driving behavior. Using K-means clustering on 19D kinematic features, four stable behavioral modes were identified: anticipatory soft, reactive closing, brake-like jerk, and an outlier category. A HistGradientBoosting classifier predicts mode membership using only the first second of the event, achieving a macro-F1 score of 0.758. Pair age was found to be the primary scene context modulator, while scene geometry and pedestrian proximity had negligible effects. These kinematic modes proved stable across medium speeds but were highly dependent on speed regime at low velocities.
Developing socially compliant navigation for autonomous mobile robots requires addressing both spatial and temporal aspects of human pedestrian interactions, which prior methods fail to fully capture. Existing Reinforcement Learning approaches rely on static reward functions, limiting realistic behavioral reproduction, while Imitation Learning lacks explicit modeling of social dynamics. The proposed framework integrates an imitation learning approach leveraging spatiotemporal dynamics for compliant navigation. A novel graph-based auxiliary network encodes the complex crowd state by paying attention to individual pedestrians' interactions. Furthermore, a navigation module incorporates predicted encoded states and uses a trajectory-level objective to minimize error accumulation, achieving superior performance on simulation and real-world data.
Non-terrestrial networks (NTNs) are crucial for enabling embodied intelligence by providing ubiquitous connectivity for remote robots. Existing memoryless protocols fail in the dynamic, resource-constrained environment of EI applications. This paper introduces the memory-native NTN (MemNTN) paradigm to overcome these limitations using long-horizon contexts. It establishes a dual-memory architecture managing both world state and historical network experience. Experiments on satellite embodied question answering prove MemNTN significantly outperforms traditional stateless methods.
Safe laparoscopic surgery requires accurate enforcement of the remote center of motion (RCM) constraint for stable field-of-view adjustment. Current control strategies lack a unified benchmark for comparing different RCM formulations and image-based visual servoing (IBVS) frameworks. Researchers present an open-source simulation framework integrating three RCM models and six IBVS control architectures in a unified velocity formulation. This platform allows controlled evaluation of how modeling choices affect performance under constraint. Structured case studies highlight structural sensitivities related to tangent-plane definition, dimensionality, and robustness near singularities.
Developing complex robotic manipulation requires abstracting continuous sensorimotor data into discrete object and action representations. The proposed model jointly discovers high-level manipulation primitives and object categories by predicting multi-modal outcomes—including motion, contact, and force—from random interaction data via a binary bottleneck layer. This allows for using the learned representations in a discrete planning method that supports partial actions for fine-grained control. Furthermore, the framework achieves few-shot generalization to novel objects by comparing limited interaction effects against predicted object symbols. Experiments on repositioning and stacking confirm that this effect-driven approach surpasses existing visual and state-of-the-art methods in planning accuracy.