Force-Sensing Robots, Not Humanoid Hype, Are the Real Physical AI Bottleneck

Robots still aren’t very good at picking things up, despite billions of dollars of investment and endless videos fed into deep-learning models. That may be about to change, thanks to a manipulation architecture built by Eka Robotics, which encompasses a force-sensitive gripper and a new kind of deep-learning model. Founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, the startup closed a $150 million Series A round valuing it at $1.45 billion post-money, led by Andreessen Horowitz and F-Prime, on October 20, 2026.

That matters because robot manipulation keeps striking a single wall: cadres of humanoid robots draw most of the venture financing, while the thorniest technical piece of robot manipulation — force sensing, the “contact” part of contact — gets raised round after round by teams of finger-flyers rather than by one overwhelming contender. Eka’s raise reframes where the bottleneck actually sits.

In Eka’s industrial framing, the question becomes tractable statistically. Picking a dish detergent bottle, a bin of loose parts, or an oddly angled engine block off an assembly line has long been solved in highly structured production where fixtures hold every part in place. Force-sensing on top turns those applications into a solvable learning problem rather than the bin-picking cliché that robots in unstructured settings have trapped themselves in.

Photo: a commercially produced five-fingered servo-electric robot hand exhibited at Hannover Messe in 2016. Eka’s bet is that a model trained on force, not just vision, is what lets machines manipulate objects with the same care — illustrated here, not Eka hardware.

Why force, not vision, is the missing sense

The most generalizable physical AI models today — the ones that can attempt lots of different tasks — are not particularly good at them. But AI models trained to be highly capable at a single task may not close the business case for a company. Eka’s foundational bet is that the contact moment is where the gap lives.

Rather than a vision-language-action model, which is trained to understand the relationship between language, visual inputs, and actions, the schema Eka uses is a vision-force-action model, which focuses on the relationship between language, force inputs, and actions. Critics of the current direction of generative AI robotics models, like Rodney Brooks, say such models will fail without an attempt to represent force and touch. Eka’s results suggest those critics have a point.

“Vision is not the primary way in which you close this last millimeter problem. Vision tells you roughly where the object is, what the pose is, but the actual grabbing, the actual searching is happening through forces.” — Pulkit Agrawal, Eka co-founder

Eka had to design a unique robotic hand that was more flexible than traditional grippers and included more force sensors. It reacts to contact, which is why it doesn’t crush a raspberry. The founders are cagey about how they achieved that for competitive reasons, but it comes down to using a physics simulator to model the multiple forces the robot will encounter as it attempts to grasp something, and using that to create a reinforcement-learning environment that can be run at scale. They say they can generate thousands of hours of robot experience in minutes.

The market is quietly shifting away from humanoid-only betting

Arthur D. Little’s analysis of physical AI investment suggests the imbalance is already correcting. Its analysts argue the market’s center of gravity is shifting toward specialized force-aware manipulation in controlled industrial settings — food-grade grippers for sushi and salmon, disposable soft grippers for pharma vials, incision-aware surgical assistants — each of which fails quietly or dangerously without an accurate force read. Orders for such systems are quoted at 30–70% higher prices than comparable forceless “smart grippers,” but with explicitly weaker guarantees on safety-critical cycles.

Whether that shift plays out fast enough to get ahead of the rest of the field is an open question. Public humanoid-robot prototypes across the sector have recently stressed the same challenge in demos rather than in papers: at least one half-billion-dollar-valued humanoid company paused a state-of-the-art field-trial event after repeated failures to lift a silicon-chip e-commerce carton at a fixed orientation. A failed grip is still a real one — and it’s easier to train a model on contact-sensor inputs that relate to it than it is to write the logic in production-grade safety code, where a miscalibrated squeeze can crack the product or crush the packaging it came in.

Eka’s backing and reported performance

Eka reports up to 1,500 picks per hour in laboratory induction and bin-picking tasks — two and a half times faster than humans — and says it reliably picks objects more than 95% of the time. The robot arm is strong enough to hurt itself as it moves that quickly. One frontier-lab observer said it had a “wow factor,” noting the robot could adapt to a large number of different objects without retraining. The hand costs less than $1,000 and is rated for 5 million duty cycles; the ultimate goal is to hit the 20 million cycle bar set by standard industrial arms.

Eka’s broader team comes from MIT, Berkeley, Harvard, CMU, BU, UPenn, DeepMind, Microsoft, and Boston Dynamics, spanning self-supervised learning, large-scale robot learning, world models, reinforcement learning, and sim-to-real robotics. Andreessen Horowitz partner Raghu Raghuram, who co-led the round, said Eka’s capabilities are “already exceeding superhuman performance for certain tasks,” with superhuman performance expected in many dexterous manipulation tasks over the coming months.

Photo: a Rethink Robotics employee teaching Baxter, an early collaborative robot arm, at an exhibition. The manual “teaching” posture mirrors Eka’s bet that physical contact itself should carry information — illustrated here, not Eka hardware.

A same-day physical AI pairing: NEURA and NXAI

Eka was not the only physical AI story on October 20. NEURA Robotics and NXAI announced a partnership the same day to jointly develop a memory component for cognitive robots, integrating NEURA’s Physical AI stack with NXAI’s xLSTM architecture. NEURA describes itself as building cognitive robots that can see, hear, feel and learn; its Neuron AI multimodal foundation model encodes 3D vision, tactile and force-torque data, audio, robot state, and language in one shared space over time. The partnership adds working and episodic memory so robots can build on experience over time rather than only execute individual skills.

The pairing is instructive because it shows the physical AI stack converging from multiple directions at once: force sensing at the gripper end (Eka), memory at the cognition end (NEURA/NXAI), and the sim-to-real bridging problem underneath both. Whether the sim-to-real gap closes fast enough to keep funding ahead of hype remains the question hanging over the sector.

What it means for the sector

Robotics suppliers rarely get the press attention of investor decks. But the math could eventually do the talking: the workload that makes the trillion-dollar-number arguments about AI possible is, ultimately, contact-rich manipulation at scale. Whoever owns the force-vision-contact layer controls the gate everyone else’s needs have to pass through — and that is where this month’s money went.

If force-sensing turns out to be the missing sense in the stack — the way text sequencing was for language — a $150 million raise at a $1.45 billion valuation will look cheap. If it turns out to be the usual addition of “yet another modality,” the valuation will end up more famous than the product line. For now, the physics of contact are doing the arguing: when a heavy arm rushes toward your hand and slows down just in time to touch gently, that is not a vision trick — it is force.

For more on how robotics is reshaping industrial automation, follow our ongoing robotics and drones coverage.

References

• TechCrunch — Eka Robotics Raises $150M at $1.45B Valuation on Vision-Force-Action Model
• Unite.ai — Eka Robotics Secures Series A Led by Andreessen Horowitz
• Unite.ai — NEURA Robotics, NXAI Team Up to Give Cognitive Robots Lasting Memory
• Wikimedia Commons images: Schunk SVH robot hand and Rethink Robotics Baxter