Robotic Lifestyle

Physical Intelligence raises $600M to advance robot foundation models - The Robot Report
Humanoids·3 min read

From Foundation Models to Digital Twins: The engineering path to safe telesurgery and humanoid manipulation

By Sophia Chen

Surgeons in a flagship hospital, a robot arm 2,000 miles away, and an AI that predicts the next 50 motion steps in 100 milliseconds: that is not science fiction; it is the techno-trajectory now being built by companies raising hundreds of millions, writing cybersecurity playbooks, and stitching simulation to real-world robots.

Three technical trends are colliding this winter: large robotics foundation models that translate vision into action, hyper-realistic digital twins that stress-test controllers, and a fresh push to enable telesurgery across regulatory regimes. Each addresses a different failure mode - perception, validation, and connectivity - that has kept advanced manipulation out of routine clinical practice.

How foundation models give robots surgical sense

Why this matters now: investors and device makers are funding pilots and standards at scale. Physical Intelligence closed a $600 million Series B in late November 2025 to push its vision-language-action models; Sovato Health, which recently attracted an Intuitive Surgical investor in a $26 million Series B, is actively advising robotic OEMs on how to make remote procedures safe; and Parallax Worlds is selling high-fidelity virtual testbeds to compress months of on-site tuning into days. Those are not incremental bets - they change what a lab prototype needs before it hits the operating room or factory floor.

Robotics foundation models move beyond hand-coded trajectories. Physical Intelligence describes a pipeline in which RGB-D streams and joint histories are tokenized and fed to a 3-5 billion-parameter transformer that outputs the next micro-actions. The company says its VLA model can predict 50 future steps in roughly 100 milliseconds, enabling tight closed-loop control at human-scale task frequency.

Digital twins: the rehearsal that reduces patient risk

That capability matters for surgery because soft-tissue manipulation depends on continuous sensory feedback and tiny corrections. In plain terms, the model is learning the physics of interaction - how a needle bends, how tissue yields - rather than merely replaying a recorded motion. Physical Intelligence reported that using RECAP (reinforcement with demonstrations and corrections) doubled throughput on delicate tasks such as filter insertion or folding, and lowered failure rates over hours of operation.

There are limits. Transformer-based policies still need safety envelopes: force thresholds, constrained speed, and fallback behaviors. For humanoid-style manipulators used in clinical settings, engineers must certify that the model cannot command motions outside certified joint limits and must demonstrate graceful degradation when sensors disagree or latency spikes.

Regulatory, latency and cybersecurity hurdles

Testing a new surgical workflow on real patients is slow and hazardous. Parallax Worlds aims to virtualize the physical space so developers can run the same surgical control software against a hyper-realistic twin. The startup claims it can turn simple phone video walkthroughs into interactive 3D environments compatible with physics engines such as NVIDIA Omniverse, then run the real robot stack inside that virtual room.

For humanoid platforms and multi-instrument OR rigs, that matters because edge cases - a wet surface, a slipping clamp, or an odd anatomical variance - are where systems fail. Parallax’s approach lets teams generate thousands of adversarial scenarios and measure metrics such as time-to-failure, mean time between critical faults, and success distribution across patient geometries before any hardware moves.

Manufacturers and hospital systems are already paying for that confidence. Parallax said it signed five robotics customers across manufacturing and construction verticals and raised a $4 million seed round in November 2025 to scale R&D. This mirrors a broader shift in industry: simulation is no longer just for design; it is now a regulatory and reliability tool.

Regulatory, latency and cybersecurity hurdles

Sources