Silica, a robotics startup backed by Y Combinator, has launched a line of torque-sensing actuators, which it says add only 10 to 15 percent to the cost of a robot joint. That figure is an unverified company target: public materials show no dollar price and no like-for-like cost comparison. The company expects to deliver prototypes in December 2026.
The public materials examined show no funding round, no public price and no independent test data, so for now the pitch rests on design targets rather than demonstrated performance.
A Crowded Field for a Self-Described First
Silica presents its actuators as the first affordable force-sensing actuator. The record does not bear out "first" as an industry milestone. HEBI Robotics, according to the Y Combinator company page and manufacturer documentation, dates its series-elastic X-Series to 2016. Sensodrive sells integrated torque-sensing drives under the SensoJoint name, and Harmonic Drive markets gears with torque sensors built in.
What may distinguish Silica is price and fit. The company says it is aiming at inexpensive robot joints, including those in low-cost training arms, and expects eventual use in factory and laboratory machines as well as humanoids.
The word "force" also needs a qualifier. On Silica's website it mostly means measuring torque at the joint output. That is a different thing from a six-axis wrist force/torque sensor or a tactile skin that resolves contact across a surface, both of which competitors and researchers treat as separate sensing approaches.
What Silica Has and Hasn't Shown
The company's purchase page lists customer prototypes for December 2026 and production for January 2027, and it labels its specifications as design targets. The Y Combinator launch copy, however, says samples ship "in February," without naming a year. The two schedules do not square, and the discrepancy is one a reporter would want the founders to explain.
Silica also reports requests for more than 60,000 units within a month of opening preorders. That number measures interest, and little more. Nothing in the public statement establishes paid orders, noncancelable commitments, delivered units or revenue, and the launch page carries no date.
Other gaps are larger. No independently reproduced noise floor, calibration protocol or bandwidth measurement has surfaced, nor has a like-for-like cost comparison. The purchase page shows no dollar price. RoboHorizon's coverage of a tenfold figure, likewise, cited no independent instrumented benchmark.
Even the company's origins are a little blurry. Dealroom pages disagree on launch timing, listing April 2026 in one place and September 2026 in another. Y Combinator confirms the founding year but not an exact incorporation date.
Why Cheaper Sensing Matters
Demand for robots in general is not in doubt. The International Federation of Robotics (IFR) has projected 655,000 industrial robot installations in 2026 and 806,000 in 2029. Those are forecasts rather than completed-year results, and they say nothing about demand for Silica's product in particular.
Humanoids face a stiffer test. The IFR says they must meet requirements including cycle time, energy use, maintenance costs, durability and consistent performance before they can compete with conventional automation. Cost sits on that list, and cost is the lever Silica is pulling.
Sensing hardware alone, though, does not close the loop. One arXiv paper reviewed for this article argues that measurement hardware is only part of the stack: policy outputs, control interfaces and training supervision must also represent force, or the data goes unused.
The Alternatives Buyers Will Weigh
Sensodrive offers the clearest point of comparison. Its manufacturer page lists SensoJoint 3000 variants with rated torque of 4 to 7.8 N·m and maximum torque of 9 to 28 N·m, along with integrated torque sensing. Those figures belong to different performance classes, and are defined differently, than Silica's peak-torque targets, so they cannot be compared directly on price.
Software is a rival of another kind. Estimating torque from motor current remains a real option and should not be written off as uniformly incapable, since results hinge on modeling, calibration and the task at hand. Add-on sensors, such as ATI's force/torque units, can also bring sensing to a robot without replacing every actuator, though they capture different quantities over different areas.
Humanoid builders are approaching the problem from yet another angle. The architecture of the hand on 1X's NEO suggests that sensing quality can come from how transparent the transmission is, not only from extra embedded sensors. It is an integrated hand, however, and not a substitute for what Silica sells.
Nothing found in the reporting ties Silica to the UMI gripper project, NVIDIA or Figure, and no such relationship should be inferred.
Safety, Scale and the December Test
Buyers will want safety evidence. A sensing capability does not show that an actuator or a robot meets a safety standard. Sensodrive markets certified safety functions separately, a reminder of the additional proof purchasers may demand. Legal obligations depend on the product, the integration and the intended use, and no component-specific determination has been made for Silica.
Pressure on price could also come from established volume manufacturers. A Schaeffler investor presentation on humanoids points in that direction, though its claims are the manufacturer's own and unaudited.
The first real test arrives in December 2026, when Silica says prototypes will ship. Independent measurements of noise and bandwidth, set against existing drives on price, will show whether the company's targeted 10 to 15 percent premium, which public materials do not yet verify, holds up.
