Generalist hypothesizes that large-scale physical pretraining creates a reusable foundation for robot behavior. The method involves short demonstrations to teach new tasks instead of requiring bespoke datasets. Models across GEN-0, GEN-1, and GEN-1.5 have achieved nearly 99% success on selected tasks via in-context physical prompting. The core question remains whether these performance gains translate to successful, long-horizon real-world deployments. The work focuses on enabling robots to perform complex physical interactions through conversational learning.
Automation in warehouses necessitates specific Personal Protective Equipment (PPE) for personnel sharing space with machinery. While no single legal checklist exists, essential items like high-visibility clothing, safety footwear, and gloves are standard practice based on risk assessments. Visibility is crucial as it helps drivers and sensors perceive people, especially in challenging conditions like low light or cold storage. High-visibility clothing must be selected based on the specific role and risk zone, adhering to standards like EN ISO 20471. A comprehensive PPE strategy must also include footwear, hand protection, and proper garment maintenance for optimal safety.
Dream Labs raised $90 million in seed funding and launched the Dream API to translate natural language and camera feeds into robot motor commands. Their Dream-1 model learns how to perform tasks by observing video, allowing it to control new robots without prior specific data. This capability enables zero-shot learning, meaning robots can perform new tasks quickly across different hardware. The Dream API allows developers to connect any robot with grippers to perform tasks instantly. Dream Labs offers fine-tuning services and a robot playground for broader developer access.
Claire interviewed Jan Petrš from EPFL about novel robot designs for construction and architecture. Petrš is a roboticist and entrepreneur co-founding Déploya Robotics, which develops deployable robotic systems for automated façade installation. His research has focused on tensegrity robotics and combining reconfigurable and soft robotics for construction applications. The interview featured insights into his work at the intersection of robotics and construction.
Scientists at ETH created bioactuators by integrating living muscle tissue from mouse cells with 3D-printed hydrogel scaffolds. They designed the muscle and scaffold as a single system using computer simulations to maximize movement. This approach resulted in actuators with visible movement, capable of jumping and moving across surfaces. The successful design highlights the importance of the interaction between living tissue and soft artificial structures. Future challenges include developing long-term nutrient supply and managing the variability of living tissue.
Generative Bionics is creating Europe's largest physical AI training facility for humanoid robots, utilizing AMD computing infrastructure for real-world training and digital twin simulation. The initial 5,000-square-meter facility will be built in Italy and operational by late 2027, as part of a planned three-site network. This factory will allow humanoid robots to generate physical data by performing tasks in replicated industrial environments. Models trained through this process, involving simulations of various conditions, are then validated on physical robots in a continuous "real-to-sim-to-real" cycle. The project aims to advance European AI and industrial sovereignty by developing this infrastructure across the continent.
Gemini 4 Argon learned a physical task from video by analyzing spatial coordinates and calculating joint movements. This learned behavior was then transferred to Apptronik’s Apollo humanoid robot. The process utilized a "Video-to-Sim-to-Real" pipeline to reduce weeks of manual simulation work to a few hours. This technique aims to lower the manual engineering required to teach robots new physical skills. The system was also tested on the Embodied Reasoning benchmark, achieving an 87 percent score.
Swedish robotics startup 120ft raised €576,000 in angel funding to develop autonomous microfactories using humanoid robots for local manufacturing. The company creates modular factories to allow consumer brands to assemble products locally in smaller batches, bypassing long international supply chains. Their system integrates humanoid robots, AI, virtual simulation, and orchestration software to convert designs into digital assembly instructions. 120ft has significantly reduced robot skill training time, aiming for just seven minutes for full assembly by 2027. The funding will support the development of their factory infrastructure and software as they move toward commercial production and future repair services.
Infineon named the Italian startup AdapTronics the winner of its 2026 Startup Challenge focused on humanoid robotics and physical AI. AdapTronics won for a technology enabling robots to handle objects without conventional grippers. Other winners included COSMOKE for tactile sensing and TECHiFAB for in-memory computing in motor control. Infineon provided hardware, software, and AI development tools to the participating companies. AdapTronics will now collaborate with Infineon to further develop and scale its winning technology.
The author developed inspection robots like Spot to allow engineers to gather data from hazardous power substations remotely, thereby reducing human exposure to danger. Robots are intended to collect evidence, but they do not replace the engineer's need to interpret findings and make final decisions. An accurate reading does not always provide sufficient context, meaning physical investigation may still be necessary to fully understand the situation. Therefore, the focus should be on using robots to provide information while retaining human expertise for judgment and necessary on-site actions. The ultimate goal is to enhance safety by outsourcing hazardous tasks while keeping skilled human oversight central to the process.
The 140th Canton Fair reflects a significant shift in robotics sourcing from prototypes to production-grade, deployable systems. International buyers now prioritize systems ready for real-world application, driven by themes of innovation, green transition, and smart manufacturing. The fair now serves as a central validation hub showcasing the entire robotics value chain, from core components to end-to-end solutions. Exhibitors showcase integrated technologies and hardware for diverse applications like humanoid robots and autonomous inspection platforms. Consequently, buyers focus less on demonstrations and more on practical details like uptime, maintenance, and regulatory compliance.
Former OpenAI science lead Kevin Weil launched Warp Sciences, an AI startup focused on automating wet lab work for drug discovery. The company trains robots to synthesize and test promising compounds identified by researchers. This initiative aims to reshore the labor-intensive work currently outsourced to China and mitigate geopolitical supply-chain risks. Human chemists will train robots to perform various lab tasks, accelerating drug discovery through 24/7 robotic operation. Warp Sciences has raised $215 million and has an $865 million valuation, led by Lightspeed.
Tesla is in the process of manufacturing humanoid robots like Optimus while simultaneously working to ensure they are useful, reliable, and durable. CEO Elon Musk acknowledged that significant work remains on the robot's hands, safety, reliability, durability, and the necessary artificial intelligence. Musk anticipates that full delivery to outside companies will occur around the end of 2027, pending the resolution of these challenges. To achieve this, Tesla needs to build a large fleet of robots to generate training data through self-play. Experts suggest that solving the challenges of dexterity and reliability may take longer than Musk's timeline.
Neura Robotics unveiled the lighter 4NE1 Gen 4 humanoid robot along with a new €17 million robot training center at the Technical University of Munich (TUM). The new model features a slimmer body, upgraded hands, and Smart Limbs that enable independent environmental responses and more precise gripping. The robot operates using Neuron AI, a model that integrates vision, touch, force, and state information to guide actions. The training center facilitates practical industrial task testing for companies across various sectors. This facility links real-world tasks with data from robot movements, human demonstrations, and simulations to develop robot skills.
The G1+ is a compact humanoid robot from Unitree Robotics designed for research and education, featuring locomotion and basic manipulation. It scores highly in navigation (5/5) but has limited manipulation capabilities (1/5), demonstrated by tasks like obstacle avoidance and sorting goods. The robot measures 134 cm tall and weighs 35 kilograms, possessing abilities such as running, jumping, and walking stairs. This version includes upgrades like a stronger neck, dual cameras, and improved audio systems. The platform is available for education with SDK support, emphasizing its strength in movement studies.
4NE1 Gen 3.5 is a full-size humanoid robot from NEURA Robotics designed for industrial applications like manufacturing and logistics. It measures 180 cm tall and weighs 80 kg, featuring 55 degrees of freedom and specialized sensors including a patented sensor skin and 3D vision. The robot possesses capabilities for obstacle avoidance, sorting goods, and product assembly, with navigation rated 1/5 and manipulation rated 2/5. It is being developed with NVIDIA Jetson Thor and NEURA's electric actuators, with production expected by the end of 2026 at a listed price of USD 115,000.
The RH56H1 is a five-fingered robotic hand from Inspire Robotics, designed with a human hand format and a metal internal structure. It weighs 0.48 kilograms and measures 75 by 159 millimeters, making it suitable for humanoid robotics and prosthetics research. The hand has a low overall Hand Score of 3/10, reflecting a modest dexterity rating and lower strength rating. Research teams should consider specific application requirements when evaluating its use for various projects. The hand is currently in production with a listed price of $7,000 USD.
Walker Tienkung is a 170 cm tall, 55 kg humanoid robot designed for research and education. It possesses mobility skills including obstacle avoidance, 100m running, and stair climbing, emphasizing movement over advanced manipulation. The platform features 49 degrees of freedom, aerospace-grade alloys, and joint modules with 300 Nm peak torque. It supports software environments like Linux and ROS 2, enabling experimentation with humanoid movement and basic object handling. The robot is commercially available at USD 41,000 for academic and research use.
Phoenix Gen 8 is Sanctuary AI’s eighth generation general-purpose humanoid, featuring a humanlike upper body mounted on a wheeled base for stable positioning. It measures 170 cm tall, weighs 70 kg, and has a lifting capacity of 25 kg, utilizing electric actuators and hydraulic hands with fine tactile sensing. The robot demonstrates capabilities in obstacle avoidance, sorting goods, and product assembly, which aid in collecting data for the Carbon AI control system. The system emphasizes physical manipulation tasks over complex navigation skills, reflecting its current developmental stage. Released in December 2024, it is currently listed as a prototype with a price of USD 150,000.
Figure's roadmap has completely changed over the past year, according to Brett Adcock. The team size has also doubled in response to this progress. Adcock indicated that the changes reflect rapid progress rather than setbacks. He shared an internal staff meeting publicly to align employees on the company's mission. This update provided an overview of the organizational growth at the humanoid robotics developer.
Dex is an industrial humanoid from Richtech Robotics designed for manufacturing, logistics, and warehouse environments, featuring two arms on either a mobile wheeled base or a stationary configuration. It stands 137 cm tall, weighs 60 kg, and utilizes NVIDIA Jetson Thor for its computing platform and electric motors for movement. The robot supports functions like obstacle avoidance, sorting goods, and product assembly through modular end effectors. Its current skill score of 3 out of 10 and lower navigation/manipulation ratings indicate its industrial capabilities are still developing. As a prototype priced at USD 60,000, focused trials are recommended before production deployment.
The Advanced Robotics Conference 2026 will be held in Aachen on November 25–26, 2026, focusing on Physical AI, particularly in humanoid robotics. The event will feature 19 speakers covering topics from AI software and hardware to sensors, actuators, and industrial applications. A central theme is addressing the challenges of deploying reliable and safe robotic systems in real-world industrial settings. Discussions will encompass the integration of various technologies and the economic viability of advanced robotics. The conference aims to connect research, engineering, and industrial applications to translate technological advancements into practical solutions.
SlipSense uses a multimodal tactile sensor, the TacV5, combining a piezoresistive array and an accelerometer to detect object slipping with high accuracy. The system detects slip events within 23.1 milliseconds, achieving a 96.7% Macro F1 score on a large dataset of 1.4 million frames. The framework utilizes modality-specific encoders and cross-attention to fuse the spatial pressure and vibration cues, demonstrating that the two modalities are crucially complementary. This system achieves zero-shot transfer to new robotic platforms without retraining, showcasing strong generalization. Touch2Trace further applies tactile feedback to dexterous cable tracing, showing that this tactile input significantly improves performance over proprioception-only methods.
The TerraTransfer++ method proposes a demonstration-free, end-to-end planning approach for autonomous driving. It leverages self-play in vectorized simulators to generate a diverse distribution of driving scenarios without relying on expert demonstrations. A flow-matching trajectory teacher is distilled from this self-play data to guide the planning process. Lightweight adapters are trained to reproduce the teacher's latent space using specific loss functions. This method yields superior closed-loop policies and planners compared to imitation-based methods on relevant benchmarks.
Designing effective reward functions for humanoid locomotion in reinforcement learning is challenging due to issues with naive reward shaping. Multi-Potential-Based Reward Shaping (MPBRS) is proposed to address this by shaping rewards using potential function differences to guide exploration without altering the optimal policy. The study implements MPBRS for humanoid locomotion and demonstrates a unified reward configuration across nine different robots. Experiments show that MPBRS yields more consistent scaling and reduced sensitivity to tuning compared to standard reward shaping. This makes tuning reward functions simpler across various humanoid robots.
Tactile sensing is crucial for dexterous robotic manipulation by providing direct access to physical contact. Research in this area has evolved from mechanics-driven methods, which interpret signals through contact mechanics, to data-driven approaches learning from tactile data. The paper reviews the fundamentals of tactile sensing and the progression from mechanics-driven to data-driven manipulation. Data-driven methods include imitation learning, reinforcement learning, and vision-language-action models for tactile control. Open challenges remain in achieving scalable, generalizable, and robust tactile-enabled robotic systems.
A sensing and control system was developed for autonomous needle insertion within a CT bore robotic arm. This system integrates respiratory sensing to achieve gated needle insertion, eliminating the need for manual patient breath-holding. Experiments confirmed the system's capability to track and predict patient breath cycles and detect anomalies. The goal is to automate CT-guided lung biopsy to reduce procedure time and patient discomfort. This approach aims to decrease radiation exposure during the medical procedure.
Predicting lane-change intentions allows autonomous vehicles to make proactive decisions instead of reactive ones. The study evaluates three driving strategies—reactive, prediction-based longitudinal action, and prediction-based lateral action—in simulated merging scenarios. Prediction-based strategies successfully avoided collisions, reducing peak deceleration before impact by 31.5% and 91.0% compared to reactive driving. Ride comfort is assessed using ISO 2631-1-based horizontal vibration estimates. Overall, lane-change prediction supports safer, smoother driving by enabling proactive decision-making and ride comfort evaluation.