Robot Manipulation

Mega Series A rounds are the dominant capital pattern

Series A rounds make up 5 of the 10 deals in the last 90 days and about $885M of the total, while seed and Series B are nearly empty ($13M). The week of 2026-08-31 saw a $500M Series A from Andreessen Horowitz and Accel. The week of 2026-10-19 saw a $150M Series A at a $1.45B valuation from a16z and F-Prime, reported by both StrictlyVC and Axios Pro Rata. An $85M Series A followed in early September. Axios Pro Rata notes that large checks are going to robotics and physical AI at early stages, so the scarce resource is conviction in a few teams rather than broad deal flow.

a16z and Andreessen Horowitz appear in 4 deals each and are the top investor by count. Mind Robotics, carved out of Rivian by RJ Scaringe with Eclipse, will do assembly work on Rivian factory floors with high-ROI use cases while its models learn from the work. Ajay Agarwal argues that robotics winners need vertical integration, citing Kiva's acquisition by Amazon. Captive deployment inside a parent company is becoming the template.

Why it matters · Winners may be those with a built-in customer and a data loop, not those selling general-purpose arms.

Research papers on physical AI keep using the same hardware. These include Franka Panda and FR3, the UR5e and UR5 from Universal Robots, UFactory xArm6 and xArm7, and the Robotiq 2F-85 gripper. Cross-embodiment transfer lets a policy trained on a Franka deploy on a UR5. Universal Robots' UR AI Trainer is described as an emerging standard bimanual setup.

Why it matters · The hardware layer is standardizing, so differentiation is moving to models, data and software rather than arms.

Policies trained in simulation are now reaching real hardware. One reached 73% success (173 of 237 trials) on a Flexiv Rizon 4S with a Robotiq gripper, and a distilled RGB policy reached 94% on gear-mesh insertion on a UR5e. A VLA that outputs a stiffness target alongside pose, and a policy that varies wipe force from 6.4N to 9.1N on language instruction, show force and tactile feedback becoming first-class. ForceTwin on the Franka FR3 reached 81.2% completion against 1.6% for a VLM prior.

Why it matters · Industrial assembly is becoming addressable without demonstrations, which favors force-control and tactile hardware vendors such as Flexiv, GelSight and Contactile.

Sharpa's 22-DoF Wave Hand is being used on Franka Research 3 arms and on bimanual stations, and a new paper introduces in-hand assembly as a manipulation setting. Cross-morphology transfer from parallel grippers to dexterous hands is also being studied. Flexiv's compliant Rizon 4 arm is used as a reactive base for a dexterous hand.

Why it matters · Hand makers are becoming a distinct hardware layer, and capability is shifting from pick-and-place toward multi-object dexterity.