OpenAI's Decisions API Is Now Generally Available. At $0.10 Per Million Tokens and 10x Faster Than Standard LLM Calls, It Creates a Dedicated Decision Layer for Enterprise AI Agents.

In October 2026, most robotics companies are still trying to teach robots to see. Eka Robotics is teaching them to feel. The difference explains why the startup, founded in 2025 with 36 employees, raised a $150 million Series A at a $1.45 billion valuation — and why investors who have watched dozens of robotics cycles say this one is structurally different.

The company was co-founded by Pulkit Agrawal, a professor at MIT's Computer Science and Artificial Intelligence Laboratory, and Tuomas Haarnoja, the former Google DeepMind researcher who invented soft actor-critic — the reinforcement learning algorithm underlying most modern robotics foundation model training. Their conviction is that the robotics industry has been optimizing the wrong modality. Vision is necessary but not sufficient. To handle objects the way a human does — picking up a strawberry without bruising it, seating a GPU memory module into a server rack without bending the pins — a robot needs to feel resistance, pressure, and surface texture in real time.

The result is what Eka calls the Vision-Force-Action (VFA) model: a foundation model architecture that takes both camera images and live force sensor data as inputs, and returns motor control commands as outputs. On the company's benchmark tasks, the model handles approximately 1,500 objects per hour at a success rate above 95% — compared to roughly 600 objects per hour for a skilled human packer. Andreessen Horowitz's Raghu Raguram called it "superhuman performance on certain tasks." Duke University's Kris Hauser, an outside robotics professor, said it could handle novel objects without retraining — the property that separates a demo from a product.

This is not a story about a better robot arm. It is a story about what happens when a foundation model research team treats haptic sensing as a first-class input — and what the $150 million raised at $1.45 billion tells us about where physical AI investment is going.

The Manipulation Problem That Stymied Robotics for Forty Years

Robotic manipulation — the ability to reliably pick up, orient, assemble, and place arbitrary objects — has been one of the hardest unsolved problems in industrial automation since the first robot arms appeared in manufacturing in the 1960s. The core difficulty is not mechanical. Robot arms are extraordinarily precise for repetitive, pre-programmed motions. The problem is perception and adaptation.

A human picking strawberries into a packing box does something extraordinarily complex without thinking about it. They adjust grip pressure continuously based on haptic feedback — tightening when a berry is slipping, loosening when resistance increases as the berry seats into place. They recognize a berry with a mold spot and re-sort it. They handle berries irregular in size, shape, and surface texture without any prior knowledge of those specific berries. They do this at roughly 600 items per hour for a skilled packer.

Traditional industrial robots handle this by avoiding it. They operate on rigid, pre-specified objects in controlled environments where position and orientation are fixed by fixtures and feeders. Any variation — a part presented at the wrong angle, an object with an unfamiliar surface, a slight change in part geometry — causes either a failure or a manual intervention. The integration cost of these systems frequently exceeds the cost of the robot itself.

The vision-language-action (VLA) model approach, which dominates current robotics foundation model research, addresses this by adding language and visual instruction following to robot control. Models like RT-X, OpenVLA, and Pi-0 generalize across task types and object categories in ways rigid industrial programs cannot. But they still rely primarily on visual perception — and vision alone cannot distinguish between a firm berry and a bruised one, or tell whether a GPU module is properly seated without the contact force rising to the expected value.

This is the gap Eka is attacking. And it has been unsolved for four decades not because nobody knew force sensing was necessary, but because the cost of adding it at industrial scale made the economics prohibitive.

What Vision-Force-Action Does Differently

The VFA architecture adds a third input modality to the standard vision-action pipeline. A custom force-sensitive gripper — engineered by Eka and priced at under $1,000 per unit — contains force and torque sensors that report grip firmness, contact area, and directional force vectors at high frequency. This data stream feeds into the model in real time alongside the camera image stream.

The practical results, according to a16z's investment announcement, include capabilities that vision-only systems cannot replicate:

• Correcting a slipping grasp mid-motion — detecting the change in force distribution before the object has moved enough to appear visually different
• Using surface contact to locate objects when visual uncertainty is high — finding a component in a cluttered bin by feel rather than entirely by sight
• Allowing insertion resistance to alter motion rather than throwing an error when a part does not seat as expected — the model adjusts trajectory in response to force feedback
• Distinguishing fragile objects by compliance — a bruised strawberry deforms differently under identical grip force than a firm one, and the force sensor detects the difference where a camera cannot

The model is trained primarily in simulation using reinforcement learning. Haarnoja's soft actor-critic algorithm has made robotic reinforcement learning dramatically more sample-efficient. Eka's simulation stack includes a high-fidelity physical model of the force-sensing gripper, enabling force data from simulation to transfer reliably to hardware — the traditional failure point for sim-to-real transfer in dexterous manipulation.

The $1,000 Gripper as the Real Moat

Force-sensitive end effectors have existed in industrial robotics for decades. What has historically made them economically impractical for general-purpose use is cost: specialized tactile sensors from suppliers like ATI Industrial Automation or Robotiq typically price between $5,000 and $30,000 per unit. At those prices, warehouse automation deploying arrays of dozens of robots faces prohibitive gripper hardware costs.

Eka's engineering decision to co-design the gripper and the model — rather than adapting an existing sensor to a separately developed model — allowed it to optimize sensor design specifically for the features the model uses. The result is a custom force-sensitive gripper at under $1,000 per unit.

The competitive implication of the co-design is critical. A larger robotics company adding force sensing to a VLA model trained without it cannot simply swap in a force sensor — it requires retraining from scratch on data the sensor produces. The architectural coupling between Eka's model and its gripper means incumbents cannot match the capability through a software update. They would need years of new training data from a new sensor, starting from scratch on a capability Eka will have been refining in production for those same years.

Performance Numbers: What 1,500 Objects Per Hour Actually Means

Eka's reported benchmark figures — 1,500 objects per hour at above-95% success rate — are company-reported on Eka's own benchmark tasks, not independently verified on a standardized test set. The appropriate response is not to dismiss them but to understand what they would mean if confirmed.

At 1,500 picks per hour, the system operates at roughly 2.5 times the throughput of a skilled human packer on comparable tasks. The more important number is success rate on novel objects — objects the robot has not seen during training — because that is where vision-only systems degrade badly and where force sensing provides the most leverage.

The outside validation from Duke's Kris Hauser — that the system handles novel objects without retraining — is more significant than the throughput numbers. In real warehouse environments, SKU proliferation and seasonal variation mean the object set changes constantly. A system that requires retraining on new objects requires expensive roboticist time and a deployment pause; a system that generalizes through haptic sensing handles variation as a feature, not a problem.

The Founders and the Research Lineage

Pulkit Agrawal runs MIT's Improbable AI Lab, with research focused on robot learning in unstructured environments — terrain traversal for legged robots, dexterous hand control, and manipulation skill transfer between robots. Tuomas Haarnoja invented soft actor-critic at UC Berkeley and spent years at Google DeepMind working on robotic manipulation and legged locomotion.

The combination is unusual: Agrawal's research specialty is exactly the domain where force sensing matters most, and Haarnoja brings the RL infrastructure to train at scale. Seed backing from MIT's E14 Fund signals that the research-to-product transfer pipeline was credible to people who saw the underlying work before it became a pitch deck.

Signal's analysis of the robotics mega-round era documented that the investors writing the largest checks in physical AI are evaluating foundation model architectures with the same rigor they brought to large language model investments in 2022–2023. Eka's founding team is precisely the kind of research pedigree that wins that evaluation.

Why a16z Led This Round

Andreessen Horowitz's Raghu Raguram stated that the firm believes "solving dexterous manipulation is one of the largest opportunities in robotics" and that "much of the world's physical economy still depends on people using their hands." The investment thesis is specific: dexterous manipulation is the bottleneck in the physical economy that neither traditional industrial automation nor first-generation AI robots have solved.

The Bureau of Labor Statistics counts approximately 4 million hand-packing, assembly, and inspection workers in the United States alone, across food processing, electronics manufacturing, apparel, and warehousing. In most of these sectors, automation rates remain low despite decades of robotics investment precisely because existing systems cannot handle object variability encountered daily. The robotics renaissance in 2026 has been driven by foundation model architectures that generalize better than prior approaches — and Eka's bet is that the next breakthrough is sensory, not just visual.

F-Prime Capital, the venture arm of Fidelity, co-led the round alongside a16z. F-Prime manages long-duration capital and typically invests in companies with 5-10 year hardware-intensive roadmaps — a signal that the investment timeline matches the capital intensity of scaling a hardware-plus-software robotics platform.

Enterprise Use Cases: Strawberry Packing to Server Assembly

Eka has disclosed two specific use cases, and the contrast is intentional.

Food processing — packing strawberries — represents the largest single category of manipulation-intensive labor in the physical economy: food and consumer goods processing, where millions of workers perform highly repetitive, force-sensitive packing tasks, automation rates are low, margin pressure is intense, and labor availability is structurally declining. A robot that packs 1,500 berries per hour at 95% accuracy, selects out damaged fruit, and adjusts grip pressure automatically solves a real operational problem without the fixture engineering traditional automation requires.

High-precision electronics manufacturing — assembling GPU server racks — represents the opposite end of the value spectrum: high-value products, tight tolerances, and expensive consequences for assembly errors. Inserting memory modules, seating CPUs, and routing cables are tasks where force feedback is critical and where a missed connection or bent pin costs hundreds of dollars to detect and repair downstream. The force-sensing approach detects assembly problems at the point of contact rather than in post-assembly quality inspection.

Both use cases share the core structural feature: the automation task requires haptic feedback to succeed at acceptable accuracy, and vision-only systems either fail or require custom fixtures that eliminate the flexibility needed for real-world deployment.

The Competitive Landscape

The robotics foundation model space in 2026 is crowded with well-funded companies. Figure AI is at a $39 billion valuation after its BMW deployment. 1X Technologies, Skild AI, Physical Intelligence, and Apptronik have all raised nine-figure rounds. Most are building VLA models with a focus on humanoid form factor and general-purpose task following.

Eka's positioning is deliberately narrow: it is not competing for the same humanoid robot market or the same benchmark leaderboard. It is competing for the specific enterprise procurement decision in manipulation-heavy environments where haptic feedback is the differentiating capability.

The competitive risk is that larger players add force sensing to their own platforms. The defense is the co-design moat: a VLA model trained without force sensor data cannot be patched with a force sensor. Retraining from scratch on new sensor inputs would take years for a company that has committed its roadmap. Eka is building the dataset and model architecture simultaneously, giving it a training data lead that compounds over deployments.

Enterprise Deployment Playbook

1. Audit manipulation tasks by force sensitivity requirement. Map every manipulation task in the target workflow and classify by object variability, fragility requirement, and precision insertion requirement. Tasks that score high on any of these three dimensions are VFA deployment candidates.

2. Pilot on the single highest-value force-sensitive task. Identify the one task where force sensitivity most directly reduces error rates or replaces the highest-cost manual labor. Deploy on that task, establish baseline accuracy metrics — success rate per 1,000 attempts, defect detection false positive rate — and run for 30 days before expanding.

3. Instrument force sensor data for quality analytics. The force data Eka's gripper produces is itself a quality signal. Objects requiring significantly more or less force than the norm are candidates for defect flagging. Build a parallel quality analytics pipeline that ingests force sensor data and flags anomalies — a quality inspection function alongside the automation function.

4. Budget gripper replacement as a consumable cost. The sub-$1,000 gripper is a consumable at deployment scale. Plan gripper replacement as a regular line item and track the replacement rate as a key operating metric alongside throughput and accuracy.

5. Run a three-week parallel operation against human baseline before calculating ROI. Validate success rate claims on your specific objects under your specific operating conditions. Run a parallel operation where the robot and a human worker handle identical object streams under identical conditions. The parallel operation data is what procurement committees need to justify capital expenditure.

What This Signals About the Physical AI Market

Signal's analysis of the Prometheus physical AI raise noted that the largest AI infrastructure bets in 2026 are moving from the data center to the factory floor. Eka's $1.45 billion Series A valuation at 36 employees is the most extreme expression of that thesis yet: the value the market assigns is not to current revenue but to the size of the manipulation problem that remains unsolved in the physical economy and the quality of the research team positioned to solve it.

The $150 million will fund roughly 18–24 months of development — scaling the model to more task types, expanding the gripper hardware to new geometries, and accelerating customer deployments. At each milestone, the company will either demonstrate that VFA generalizes to more deployment use cases beyond food and electronics, or need to defend the architecture's breadth against VLA systems that have added force sensing.

The enterprise infrastructure bet here is structurally similar to bets made on frontier model startups in 2022 and 2023: the technical hypothesis is specific, the founding team is among the handful of people globally most qualified to test it, and the market if correct is enormous. That combination is what a $1.45 billion Series A valuation at 36 employees is pricing.

Takeaway

Takeaway: Eka Robotics' $150 million Series A at a $1.45 billion valuation is the clearest signal yet that force sensing — long recognized as necessary for general-purpose robotic manipulation and long excluded on cost grounds — has reached production viability. The Vision-Force-Action model, co-designed with a force-sensitive gripper priced under $1,000, delivers 1,500-object-per-hour throughput at above-95% accuracy on manipulation tasks that have resisted automation for decades. For enterprise buyers in food processing, electronics manufacturing, and warehouse operations, the deployment question is no longer whether robotics can match human performance on manipulation tasks — it is which architecture best matches the force-sensitivity requirements of the specific tasks being automated. Eka's bet is that force is the missing modality. At $1.45 billion, the market has priced in a reasonable probability that it is right.