Integrated actuators are becoming increasingly practical solutions for compact robotic systems due to their ability to provide sufficient torque output while occupying limited installation space. These actuators combine multiple components into a single module, including the motor, reduction mechanism, encoder, and control electronics, helping to simplify system integration and reduce assembly complexity. The CubeMars AK45-10 is an example of an integrated actuator designed specifically for humanoid robot joint applications, delivering up to 7Nm peak torque in a compact diameter of 53mm and a weight of approximately 260g.
NORD Drivesystems has expanded its NORDAC VFD family to reach powers up to 200 hp with multi-protocol Ethernet interfaces, integrated functional safety, and wireless diagnostics. The new additions include four new frame sizes, covering a performance range of 0.33 – 200.0 hp, with a compact book-size design enabling side-by-side installation without gaps. NORDAC PRO VFDs are adaptable to individual application requirements across various industries and offer high-precision control with strong overload capabilities for constant torque applications.
Kraken Technology Group Ltd. raised $175 million in Series B funding, led by Digital Transformation Capital Partners GmbH, to reach a valuation of $1 billion. The investment will accelerate Kraken's global rollout of its autonomous maritime platforms, including K3 SCOUT and K4 MANTA, for NATO and international clients. Kraken has secured several contracts with defense ministries, USSOCOM, and European partners, and plans to deploy hardened, reliable capabilities globally.
Zenoh offers a lightweight communication layer for ROS 2 that reduces latency, minimizes bandwidth usage, and simplifies distributed system communication. It's particularly suitable for resource-constrained hardware like the Raspberry Pi and systems communicating across different networks. The author plans to document installing Zenoh with ROS 2 Humble, integrating it into existing projects, comparing DDS and Zenoh in robotics scenarios, and sharing performance observations from a real robot.
Peppy is an open-source robotics framework developed by Peppy Robotics, providing an alternative to ROS 2. It aims to build and deploy AI robots and humanoids with a modular node-based architecture in a single configuration file. The framework is designed as an open-source solution for robotics applications.
Researchers created a replay-constrained simulation framework to personalize impedance controllers for powered prosthetic legs, bypassing the need to model complex human neuromuscular control mechanisms. The framework uses MuJoCo-based simulator and replays recorded hip kinematics and feedback-based ground reaction forces from individual walking data. This leads to strong simulation-to-hardware predictive validity (Pearson r=0.96--0.997) and improved biomimicry rewards by 42--59% relative to the unpersonalized baseline.
A glove-agnostic algorithmic framework leveraging hand pose information mitigates pose-related artifacts in tactile gloves without requiring modifications. The approach augments force estimation pipelines with a residual prediction branch accounting for sensor deformations caused by hand pose changes. This method reduces the minimum detectable force (MDF) by 10.4%, 12.2%, and 18.3% across three glove designs and 15 users, providing a practical path to improving tactile gloves' usability in robotic applications.
An end-to-end, fully AMD-accelerated technology stack for embodied manipulation is presented, spanning data-center training silicon, Radeon PRO simulation/rendering GPUs, and Ryzen AI edge compute. The ROCm software stack unifies this technology stack. Four progressive demonstrations showcase the capabilities of VLA-based manipulation policies, including a Sim-to-Real manipulation pipeline trained with SmolVLA and deployed on a physical Franka arm.
Researchers Yuzhen Chen and KC Zhou present the World-Cognition Model (WCM), a human-centered embodied agent that uses the SLAK architecture to facilitate generalizable human-robot interaction. WCM enables users to interactively teach robots difficult tasks through chain-of-thought supervision, achieving an average success rate of 73.8% across nine real-world tasks. The model's asynchronous runtime allows reasoning, dialogue, and execution to proceed concurrently.
Researchers investigated the relationship between data and precision in robotic assembly, finding that the required number of demonstrations grows super-exponentially as target precision approaches a limit. A novel scaling law captures this relationship: log(N) ∝ 1/(P-c), where N is the number of demonstrations and P is the target precision. The limit precision c is not a static physical constant but an emergent property of the entire agent system, including its sensors and expert policy.
Researchers at Tokyo Institute of Technology proposed a two-stage framework to reduce response latency in dialogue robots by generating prefatory responses before speech recognition. This approach, called context-aware preface generation, was evaluated in a field experiment with a route-guidance robot and found to significantly reduce initial response latency compared to using fixed fillers or no filler at all. The results showed a trade-off between reducing initial response latency and increasing the gap between initial and main responses.
Wonjae Kang and three co-authors propose Sling2Sim2Real, a one-shot elastic system identification framework for non-destructive slingshot policy learning. The framework consists of two stages: (1) a multi-start Real2Sim system identification method that exploits parameter covariance to estimate elastic properties, and (2) simulation-based policy learning followed by zero-shot Sim2Real transfer using the calibrated simulator. Sling2Sim2Real reduces real-world interaction while achieving accurate policy learning and robust generalization on a slingshot manipulation task.
A novel SLAM framework, Semantic Semi-Incremental Data-Association-Free Object SLAM, jointly estimates data associations, robot poses, landmark positions, and landmark semantics from odometry and measurements. The framework leverages deep learning to create synergy between data association and landmark semantics estimation, adopts a semi-incremental estimation scheme for improved accuracy, and provides guidelines for landmark-number estimation. It is evaluated on synthetic and real-world datasets with superior performance compared to strong baselines using class labels and real-valued feature vectors.
PRISM introduces a policy representation that explicitly models polynomial interactions among observable physical variables, allowing robots to learn and represent complex interaction structures efficiently. It achieves this by applying a factorized polynomial module to the standard MLP backbone in reinforcement learning or replacing linear proprioceptive conditioning with a polynomial layer in imitation learning. PRISM improves performance over standard MLP policies and larger MLPs in humanoid locomotion and contact-rich manipulation tasks, demonstrating that interaction structure cannot be replaced by capacity alone.