Eka's $1.45B Bet: Force Feedback Meets Language-Conditioned Manipulation
Eka Robotics has closed a $150 million Series A at a $1.45 billion post-money valuation. TechCrunch
The round was led by a16z and F-Prime. The company was founded by Pulkit Agrawal, an MIT professor, and Tuomas Haarnoja, a former Google DeepMind researcher.
Eka's system pairs a force-sensitive gripper with a new deep-learning model. The company describes the model as a vision-force-action model focused on the relationship between language, force inputs, and actions. For practitioners, the ordering matters. Vision and language condition the policy, force closes the loop at contact.
The throughput claim is aggressive. Eka says the robot can pick up about 1,500 objects weighing up to 1 kg per hour, compared with about 600 for a human, with successful picks more than 95% of the time. Those figures, if sustained outside a curated demo cell, would put the system well above manual pick rates for small-parcel sorting.
Independent reaction has centered on generalization and recovery. University of Illinois Urbana-Champaign professor Kris Hauser said the demo had a "wow factor" compared to other recent robot demos. He said the robot was able to adapt to a large number of different objects without retraining. He described its persistence after a missed grasp, trying repeatedly until it secured the object, as humanness.
Force is the difference. Most vision-language-action stacks still treat contact as an afterthought, inferred from vision or from binary grasp success. Eka is making force a first-class input alongside vision and language, which directly addresses the failure modes that dominate bin picking: slip, wedging, deformation, and uncertain mass distribution.
The broader context here is a funding market that has already priced physical AI as a platform layer. SiMa.ai reached a $1.45 billion valuation after a $150 million funding round to scale physical AI, taking its total funding to $500 million as it expands its purpose-built Physical AI platform. Intelligent CIO Physical Intelligence previously raised $600 million at a $5.6 billion valuation. Bloomberg Skild AI raised $1.4 billion to develop a brain usable in robots of all shapes and sizes. Bloomberg Video Arm reorganized to create a Physical AI unit to expand its presence in the robotics market. Reuters
Looking at what this means for deployment, the questions shift from demo performance to systems issues. Throughput per hour is only useful with distributions attached: object mix, occlusion rate, exception rate, mean time between interventions, and degradation under lighting or packaging shifts. A greater than 95% pick success rate still leaves dozens of failures per hour at 1,500 picks per hour, which makes retry behavior and fault signaling as important as the nominal policy.
In this author's view, the retry detail is the most telling part of the report. Single-attempt grasp success has long been the benchmarked metric. Production value lives in autonomous recovery. A system that detects a failed or marginal grasp from force transients and re-attempts without a human reset or a full re-plan reduces the costliest part of warehouse automation, the exception path.
Worth flagging for technical readers is the language component. Language conditioning in manipulation is often dismissed as demo scripting. Tied to force feedback, it has a clearer function: specifying compliance, fragility constraints, and target affordances that cannot be read from pixels alone. The test will be whether that conditioning holds across novel SKUs without retraining, which is exactly what Hauser pointed to.
If Eka's numbers transfer, the near-term effect is not humanoid replacement. It is higher utilization of existing sortation, kitting, and returns lines where pick rate and tolerance for shape variance set output. That is less dramatic than the usual robotics narrative, and more valuable.
