A robot that only “sees” still struggles the moment an object behaves unexpectedly—slippery surfaces, soft deformations, or subtle slips. The next wave of Physical AI is built to close that gap by pairing perception with contact: visual cues tell the robot what it’s looking at, and tactile feedback tells it what’s actually happening in its fingers.
Two developments highlighted in late-September updates make the direction clear. First, researchers at Tohoku University describe a multimodal Physical AI method that combines sight and touch to manipulate objects with higher task success than approaches relying on vision alone. Second, Google’s Gemini Robotics-ER 1.5—available as a preview via Google AI Studio and the Gemini API—frames embodied reasoning as a “high-level” model capability for robots that must perceive, plan, think, use tools, and act in physical environments.
From visual perception to tactile truth
Touch is not a “nice-to-have” sensor for physical tasks; it’s the correction mechanism when the world stops matching the robot’s assumptions. Vision can identify an object’s shape and apparent pose, but it can’t reliably predict the micro-interactions that decide whether a grasp sticks: friction variations, surface coatings, minor misalignment, and how an object compresses under load.
The Tohoku University work points to this exact failure mode by integrating visual and tactile sensing into a single multimodal Physical AI approach. The reported outcome is not just better trajectories or smoother motions—it’s higher task success, which is the metric that matters in manipulation. In practice, “success” tends to hinge on whether contact leads to the intended effect (secure grasp, stable lift, correct placement) rather than merely whether the robot looks like it’s doing the right thing.
That distinction matters for robotics engineering. Systems that optimize only for visually inferred states can overfit to what cameras can measure. When tactile feedback is brought into the loop, the controller can detect whether contact is real and productive—whether the fingers are sliding, whether the grip force is sufficient, or whether the object is behaving like a rigid body versus something compliant.
Embodied reasoning: planning that survives contact
Google’s Gemini Robotics-ER 1.5 announcement reframes Physical AI as more than multimodal perception. The model is positioned as a high-level “embodied reasoning” capability—designed to help robots not only perceive and act, but also plan and use tools inside physical constraints.
The key shift is agentic planning under uncertainty. Physical environments are full of hidden variables: object weight distribution, grasp friction, and the time-delayed consequences of motion. A purely reactive system might correct its course after failure, but embodied reasoning targets a more proactive strategy—choosing actions that are robust to ambiguity and likely to work after contact.
Because Gemini Robotics-ER 1.5 is delivered as a developer-facing preview through Google AI Studio / Gemini API (announced on September 25, 2025), it signals that “reasoning for robots” is moving from research demos toward accessible tooling. This affects how quickly teams can iterate on real-world systems: higher-level decision models can reduce the amount of hand-coded state machines needed for every new manipulation scenario.
Why multimodality and agentic planning are converging
Touch + embodied reasoning is not two independent upgrades; it’s a structural change in how robots close the loop. Consider a common manipulation pipeline: a robot uses vision to estimate object pose, chooses a grasp, closes the fingers, then checks whether the object moved as expected. Without tactile sensing, the check often relies on vision again—meaning the system may only realize “it slipped” after the fact.
With tactile input, the robot can detect slip or insufficient contact at the moment it matters, and embodied reasoning can revise its plan accordingly. That closes the latency problem: the system doesn’t need to wait for the next camera frame to infer failure. The result is a higher success rate on tasks that traditionally punish robots for small perception errors.
Look at the complementary strengths implied by the two updates. Tohoku University’s contribution emphasizes multimodal sensing for manipulation success—practical competence under real contact dynamics. Google DeepMind’s update emphasizes agentic planning and tool use for operating in physical environments—scaling decision-making beyond narrow reflexes. Together, they point toward robots that can both feel and reason, with tactile feedback serving as the grounded signal that validates or invalidates a plan.
What developers should build next
The most actionable takeaway is that Physical AI systems should be designed around task success and contact-awareness, not around camera-only benchmarks. If you’re integrating tactile sensing, treat it as a first-class input to the robot’s decision loop: map tactile patterns to event-level outcomes like “secure contact” versus “slip,” and feed those outcomes back into planning rather than simply logging them.
Second, when experimenting with Gemini Robotics-ER 1.5’s embodied reasoning preview, structure your robot stack so the high-level model can request or choose lower-level behaviors—but still ground those behaviors in real sensor feedback. In other words: let the agent decide “what to do next,” but let tactile and proprioceptive signals determine whether the world is behaving as expected.
Third, run evaluation protocols that reflect physical reality. For manipulation, that typically means measuring success under deliberate perturbations—slippery variants, slight pose offsets, and objects with different compliance—because that’s where touch and robust planning pay off. The Tohoku University emphasis on higher task success suggests the right mindset: optimize for the end-to-end outcome, not intermediate accuracy.
Physical AI is moving quickly, but the direction is consistent: sight tells a robot where it is; touch tells it what’s actually happening. When embodied reasoning is built to incorporate those signals, robots stop acting like observers and start behaving like operators—agents that can recover from contact and finish the job.
