-- Atomathic, formerly Neural Propulsion Systems and a pioneer in physical AI-sensing technology, today announced a new white paper “Physical AI 2.0” that outlines a necessary architectural shift as AI moves from software workflows into embodied systems.
As AI systems move into applications, such as autonomous vehicles, robotics, and intelligent infrastructure, today’s dominant approaches—centered on vision, simulation, and reasoning—leave a critical gap that can be resolved in Physical AI 2.0.
“AI systems can only reason as well as the reality they perceive,” said Dr. Behrooz Rezvani, founder and CEO of Atomathic and author of the paper. “When observations are sparse, noisy, or physically ambiguous, the bottleneck isn’t just predicting the next scene—it is recovering the actual physical state before higher-level reasoning acts on it. Without a trustworthy state estimate, all downstream reasoning inherits upstream ambiguity.”
Defining Physical AI 2.0
While Physical AI 1.0 has been defined by large-scale simulation, synthetic data, and world models—exemplified by platforms like NVIDIA Cosmos—these "vision-first" systems are indispensable but incomplete. Physical AI 2.0 is defined by a stronger architectural sequence: world models, physical state recovery, reasoning systems, and action.
Key pillars of the Physical AI 2.0 framework include:
- World Models: Learned priors and simulation engines that generate plausible future states to expand scenario coverage.
- Physical State Recovery: The “missing link” that recovers meaningful physical structure from sparse, noisy, or degraded measurements before higher-level control begins.
- Reasoning Systems: Models that deliberate at inference-time to compare alternatives and interpret intent based on the recovered state.
- Action: Bounded actuation and safety logic that closes the loop in the physical world, generating new observations.
The Recovery Thesis
A critical vulnerability in current end-to-end models is if the observation basis is fundamentally limited—such as targets disappearing behind clutter or geometry creating blind regions—simply adding more compute or "reasoning" cannot solve the problem.
Atomathic’s approach complements the broader Physical AI stack by addressing this foundational problem of estimating the actual world when measurements are structurally ambiguous. This modular, physics-based recovery layer can exploit richer sensing and transfer across different downstream models and embodiments, from humanoid robots folding laundry to autonomous vehicles navigating long-tail road scenes.
About Atomathic
Atomathic is a pioneering physical AI-sensing technology company transforming how machines perceive and interpret complex signals in the real world. Leveraging deep expertise in advanced mathematics and proprietary AI platforms—including AIDAR™ for detection and ranging, and AISIR for Radar™ for signal intelligence reasoning—Atomathic delivers hyper-resolution sensing that enables sensors and systems to detect, interpret, and visualize ultra-high-resolution signals in real time. Atomathic’s technology is hardware-agnostic and applicable across automotive, aviation, defense, robotics, and semiconductor markets. By grounding inference in physical principles and scalable compute, Atomathic helps enable safer autonomous decision-making and intelligent machines. Atomathic can be found on the Web and LinkedIn.
###
All product and company names may be trademarks or registered trademarks of their respective holders. Use of them does not imply any affiliation with or endorsement by them.
Contact Info:
Name: Gary Bird
Email: Send Email
Organization: Atomathic
Website: https://atomathic.ai/
Release ID: 89196570
Google
RSS