Unitree has cumulatively produced approximately 18,000 bipedal humanoid robots as of July 2026. The company specifically excludes wheeled platforms or hybrid designs from its total count, focusing on purely bipedal models. This milestone comes after Unitree's record-breaking $904 million STAR Market IPO and puts pressure on rival AGIBOT in the competitive volume battle.
Researchers at Pixel Robotics have developed an approach to analyzing ROS bags using Large Language Models (LLMs) and a database-based design, similar to MCP-rosbags + mcp-lab. They utilize the mcap-mcp-server software with SQL queries for fast analysis and efficient behavior tree log retrieval through Nav2 BT logging extensions. An agent skill is also shared to simplify setup, providing a useful resource for the community.
The author spent 9 months working on an Embodied Agent project using ROS and VLM algorithms, building a simulation stack from scratch, integrating various packages, and creating a web application to configure the project. Despite completing the project with over 400 commits and 1000+ hours of work, the agent lacks realistic human-like understanding of its environment, relying solely on deterministic algorithms. The author believes that this is a major limitation, making the project feel useless and unsure if they should continue down this path.
Windows does not have a proper way to handle SIGINT, causing ungraceful shutdowns. The issue is with using `signal.SIGTERM` instead of `signal.CTRL_C_EVENT` on Windows, which is the correct signal to send for proper termination. A developer attempted to implement this change but encountered issues, including increased failures in CI runs and failed pull requests.
China is the country most prepared for self-driving cars, ranking first due to massive robotaxi deployment, nearuniversal 5G coverage, and advanced regulations. The US follows with Waymo completing about 500,000 paid rides weekly. China has more than 17 million robotaxi rides recorded, with Apollo Go operating fully driverless vehicles without a safety driver, and 96% of the population covered by 5G.
Humanoid robots are poised to revolutionize manufacturing with unprecedented flexibility, navigating cluttered spaces and handling diverse products. A $370 billion market is expected to emerge by 2040 if progress continues at current rates. However, technical challenges, such as coordinating dozens of articulated joints and simulating complex systems, must be overcome. Siemens Xcelerator offers a solution, integrating design, simulation, and manufacturing workflows to eliminate systemic design obstacles.
Roboticists and policymakers are developing country-specific roadmaps to advance robots' capabilities, interact with the physical world, and address societal needs. These roadmaps, led by academia, industry, and government entities, aim to guide innovation and investment in robotics. Examples include the US National Robotics Roadmap, the European Union's ADRA Strategic Research Agenda, and China's Made in China 2025 initiative, which prioritize areas such as collaborative robots, physical AI, and technological sovereignty.
A new algorithm for computing protection levels in vision-based pose estimation is presented, which provides probabilistic bounds on pose error under undetected faults. The algorithm covers all six degrees of freedom of the aircraft pose directly and analyzes the effect of measurement redundancy, pixel-level prediction uncertainty, and runway distance on protection levels. The results demonstrate tradeoffs in the protection levels on an illustrative runway example.
Insufficiently detailed operational data can undermine safety claims in autonomous vehicle software assessments, despite conservative analysis. This limitation can lead to dangerously optimistic reliability claims. The lack of detail can result from incomplete or inconsistent data, which can be problematic for assessing the robustness of reliability claims. Conservative Bayesian inference techniques are extended to investigate the impact of data fidelity on assessments.
Researchers have developed a method for human-following in pedestrian crowds, which navigates the proximity-safety balance by decomposing the task into sparse reward and independent cost constraints. This approach allows for explicit control over the trade-off between staying close to the target and navigating safely among surrounding pedestrians and obstacles. The method also incorporates prediction uncertainty of human motions to enhance safety under unpredictable conditions. Extensive experiments demonstrate its effectiveness in both in-distribution and out-of-distribution settings.
Whole-body planning for humanoids navigating confined spaces requires integrating self-collision avoidance into a reachability-constrained formulation. A three-stage framework formulates kinematic path planning directly over kinematically reachable rigid-body volumes, synthesizing volume-informed guides that guide a full-order trajectory optimizer over long horizons. This approach achieves restricted confinement ratios (C_r < 1.5) and generates feasible trajectories across complex foot and hand contacts, while a learned policy successfully tracks these plans under domain randomization.
FACT, a causal World-Action Model, predicts future video and task progress conditioned on the executed action. This allows failure rollouts to supervise action consequences, turning bad actions into valid future targets rather than being discarded. Extensive experiments show that FACT outperforms existing baselines, improves with failure data, and reduces success-biased future hallucination under bad actions.
A neural network-based teleoperation framework is proposed to address technical bottlenecks in direct vehicle control, including communication latency and uncertainty compensation. The system integrates the Wave Variable approach for passive stability under stochastic delays and an adaptive Radial Basis Function Network (RBFN) for active compensation of vehicle-specific uncertainties. Simulation benchmarks demonstrate that the RBFN achieves superior robustness against unmodeled disturbances while requiring orders of magnitude less execution time than other control methods. Hardware-in-the-loop experiments validate the system's practical feasibility, safety, and robust trajectory tracking under physical road uncertainties.
Researchers developed a real-world cooperative bimanual grasping framework that enables the synthesis of executable bimanual grasps from single-view inputs, reducing dependence on complete 3D object models. The approach includes a multimodal dataset, a Denoising Diffusion Probabilistic Model (DDPM)-based module that generates joint-level grasp configurations, and an execution strategy that integrates motion planning with online grasp refinement. Experiments demonstrated high success rates across unseen objects with varying geometries and poses using a dual-arm robot.
Researchers have developed a new stepping strategy for monopeds, utilizing reaction torques to counteract energetic losses from damping and increase pitch stabilization torque. This results in a user-specified balance between steady-state fore-aft speed and apex height. The strategy has been validated through simulation studies on a generic 5-link biped and physical experiments on the Penn Jerboa, achieving speeds of up to 1.77 m/s (8.85 leg lengths/s).
PBD-AG presents a persistent baseline-delta active graph framework for long-horizon service robots that autonomously builds structural baseline and inspects fixtures to ground hierarchical object beliefs. The framework maintains reliability-weighted object states over geometry, semantics, identity, existence, and support relations. Simulation experiments show higher aggregate coarse-fixture F1 scores than capability-matched controls, as well as stronger identity continuity and event recall. PBD-AG integrates with onboard sensing, providing a traceable world model for long-horizon robotic applications.
Researchers introduced SALT, a Semantically ALigned action Tokenizer, to improve language-conditioned control in vision-language-action models by aligning action representations with language. This approach enhances verb-specialized codes while maintaining reconstruction fidelity, leading to 71.9% average success in SimplerEnv, compared to 42.7% and 31.2% for other methods.