ROS 2 Lyrical Luth introduces Zenoh as the default middleware, targeting 20-50μm local latency, 0.5-2ms cross-network latency and 50% greater throughput across cloud-to-microcontroller systems.

Direct GPU/NPU interaction removes redundant memory copies and delivers major gains for deep-learning inference and point-cloud processing, with rosidl::Buffer emerging as a standard accelerator interface.

From navigation, recognition and grasping to multi-robot orchestration, behavior trees are becoming the standard pattern for complex embodied tasks.

The feature-complete rclrs client brings memory safety and zero-cost abstractions to safety-critical robotics while broadening the ROS 2 language ecosystem.

Natural-language instructions are decomposed into ROS 2 Action sequences, allowing robots to move from mechanical execution toward autonomous decisions.

TensorRT quantization, zero-copy memory and dynamic power balancing define the engineering path toward sub-200ms on-device inference.

Standardized pipelines from sensor capture to training datasets reduce the heavy cost of data preparation and support shared multi-robot datasets.

Isaac Sim → Lab → ROS → Jetson forms an end-to-end physical AI workflow across simulation, training and deployment.

Visual tools and AI code generation can compress standardized application development from months to days and lower the learning curve for new developers.

Models such as π0, GR00T N1 and OpenVLA show how vision-language-action systems may move from research to production, with distillation enabling smaller edge deployments.

A slow 2-5Hz planning system works with a 200Hz+ control system, combining deliberate reasoning with real-time motion.

The central question is shifting from whether humanoids can work to whether they are economically viable, led by hazardous-task replacement and companionship.

World models improve transfer without relying on large quantities of real data, while contact dynamics and soft-body physics remain difficult.

Dense touch sensing fills the last gap in world models, with force and vision fusion unlocking fine manipulation.

MPC + RL pipelines are maturing toward zero-transfer deployment for omnidirectional walking and slope climbing.

Pixel diffusion, implicit prediction and spatial intelligence continue to converge and compete, with VLA systems driving near-term adoption.

Cosmos, Open X-Embodiment and OpenLoong illustrate how shared data and open collaboration are accelerating physical-world model development.

From autonomous flight and air-ground coordination to low-altitude traffic management, ROS can support BVLOS frameworks and coordinated fleets.

MES-WMS-robot integration and virtual commissioning make digital twins an important step before industrial deployment.

Topics include ISO 26262/UL 4600 adaptation, learned-policy certification, model supply-chain security and deterministic safety layers beneath AI policies.