MH-1, a computer-operated mechanical hand

TL;DR — In 1962, Heinrich Ernst wired a motorized teleoperator arm to MIT's TX-0 computer, covered its gripper with binary touch switches, and wrote programs of the form "move down until touch, then close until pressure" — the first robot that manipulated the world by feeling it rather than replaying a taped trajectory. The idea (manipulation as sensor-conditioned feedback) was right; the substrate (roughly 10⁵ instructions/second, a few dozen on/off sensors, zero learning) was hopeless. Today, with optical tactile sensors giving ~10⁵ contact measurements per fingertip and policies that can be learned rather than hand-coded, Ernst's exact problem — dexterous manipulation of objects your own hand occludes — is the open frontier, and his touch-first framing is worth reviving deliberately.

What Ernst actually built

The state of the art in 1962 had two branches, and both were blind. Industrial robotics was record-and-playback: Unimate, installed at GM in 1961, ran through stored joint trajectories with no sensing whatsoever — if the part wasn't where the tape assumed, the robot stacked air. The other branch was teleoperation: master–slave manipulators for nuclear hot cells, where all the sensing and intelligence lived in a human operator behind leaded glass.

Ernst, working on his MIT doctorate (his advisor was Claude Shannon), took a servo-driven manipulator of the nuclear-hot-cell lineage and put a computer where the human used to be. Crucially, he did not try to give it eyes. He instrumented the hand: binary touch contacts on the fingers and jaw surfaces, pressure sensing in the grip, and (I believe) simple photodiode proximity elements — on the order of a few dozen channels total, most of them one bit each.

The real contribution is the software abstraction. Ernst wrote an interpreted language, MHI (Mechanical Hand Interpreter), whose core construct was the sensor-conditioned motion: a move command paired with a sensory predicate that terminates or redirects it. A block-stacking program was not a trajectory; it was a strategy:

• sweep the hand across the table until a finger contact fires;
• center on the object by alternating side touches;
• close until grip pressure;
• lift, move over the previous block, lower until touch;
• release.

With this, MH-1 could search a cluttered table, find blocks whose positions were unknown to it, pick them up, stack them, and drop them in a box — entirely by touch. That is a qualitatively different machine from Unimate: the program encodes a policy over sensor events, not a path. Ernst's own framing was essentially that intelligence in manipulation lives in the conditional structure, in how behavior branches on contact.

MH-1 (1962) MHI program on TX-0 "move down UNTIL touch" Servo arm + gripper Binary touch switches (~1 bit each) branch on contact

Why it could not work in 1962

Compute. The TX-0 was a fine machine for its day — transistorized, 18-bit words — but it ran on the order of 10⁵ instructions per second with a few kilowords of core memory. That is enough to poll switches and branch; it is nowhere near enough to estimate contact geometry, object pose, or friction from data, even had the sensors provided any. The control "loop" was event-driven and slow; fine force regulation was out of reach.

Sensing. A microswitch reports one bit: touching or not. MH-1's entire tactile stream was perhaps a few dozen bits per sample. Human fingertips, for comparison, have on the order of a hundred mechanoreceptors per cm² with sub-millimeter spatial discrimination. With one-bit contacts you can detect that you hit something, not what, where on the finger, or how it is slipping — so every disambiguation had to be done by moving and touching again, which is slow and brittle.

No learning. This is the deepest limit. Ernst had to hand-write the conditional structure for every contingency: block encountered edge-on, block tipping, two blocks touching. Sensor-conditioned programming scales as the product of contingencies — a combinatorial explosion that hand-coding cannot survive. Manipulation lives in a mess of contact modes, and Ernst was enumerating them by hand, in assembly-adjacent code, on a machine with kilowords of memory. He proved the framing; he could not scale it. There was no way to learn the policy: perceptrons existed, but nothing that could map raw sensor streams to motor commands, and no data to train on anyway.

What changed

Three things, one per limit.

Tactile sensing became rich. GelSight-class optical sensors (a camera watching an illuminated elastomer deform) turn a fingertip into roughly a 10⁵-pixel contact-geometry imager at tens of Hz — you get local shape, shear, incipient slip. Cheaper magnetic and capacitive skins (AnySkin, uSkin, and similar) give thousands of taxels over whole fingers. Meta's Digit 360 packs high-resolution touch plus vibration sensing into a fingertip. Ernst's one-bit switches have become dense images of contact.

Policies are learned, not enumerated. The combinatorial explosion that killed hand-coded MHI programs is exactly what deep RL and imitation learning absorb. Simulators now render contact and (approximately) tactile signals, so policies can be trained across millions of contact events and distilled to real hardware. Diffusion/transformer policies condition on multimodal streams natively.

The touch-first regime has been demonstrated. This is the striking part: Ernst's specific bet — manipulation without vision — has been vindicated in miniature. Qi and colleagues' "rotating without seeing" line of work showed in-hand object rotation on a multi-finger hand using only touch and proprioception, no camera. Meta's Sparsh showed self-supervised representation learning over raw tactile data transferring across sensors and tasks. NYU's tactile-dexterity work showed touch-primary policies for precise real-world tasks. Sixty-plus years on, the field is re-deriving MH-1's thesis with better substrate.

Meanwhile the mainstream — vision-language-action models trained on camera streams — struggles precisely where Ernst worked: the hand occludes the object at the moment of contact, cameras cannot see forces, and millimeter-scale contact events are below visual resolution and latency. Vision-first manipulation is playback with better trajectories more often than we like to admit.

What a serious 2026 revival looks like

Take Ernst's framing literally: touch is the primary observation; vision is the auxiliary that gets you into contact.

• Hardware: a 16-DoF hand (LEAP/Allegro-class) with optical tactile fingertips and taxel skin on phalanges and palm — order 10⁵–10⁶ tactile measurements per second, versus MH-1's tens of bits.
• Training: large-scale simulation with learned tactile rendering (train a network to map simulated contact states to realistic sensor images, closing the tactile sim-to-real gap), then RL or teleop imitation for tasks chosen so vision cannot solve them: retrieving a specific key from a pocket, in-hand reorientation of an unseen object in the dark, blind connector insertion, finding and turning a screw inside a housing.
• What to reuse from Ernst: the event structure. His guarded moves are, in modern terms, options terminated by contact events. A hierarchical policy whose high level composes contact-conditioned skills — "descend until contact," "slide along edge until corner" — is MHI with the branches learned instead of written. There is growing evidence that giving policies explicit contact-event structure improves sample efficiency over raw end-to-end control; Ernst's decomposition is a good prior.
• What to replace: everything about state. MH-1 had no memory of what it was touching; a modern version maintains a learned latent estimate of object pose and contact configuration, updated from tactile flow — touch-based SLAM over the object, which several groups are actively pursuing.
• The benchmark that would make the point: a suite where every task is scored with the cameras off. If a policy family wins there and remains competitive with cameras on, touch-first is the right default, not a curiosity.

Status: partially vindicated, mostly open

The lineage is real and traceable: Ernst's guarded moves were formalized at IBM and Stanford in the 1970s, became force-guarded motion primitives in industrial controllers, and every peg-in-hole insertion routine running in a factory today is a direct descendant. Touch-only in-hand rotation and tactile representation learning vindicate the blind-manipulation bet specifically.

What remains open is everything at scale: tactile sensors that survive months of contact (elastomers wear out — a problem Ernst's microswitches did not have), tactile sim-to-real that transfers as well as visual domain randomization does, any large pretraining corpus of touch (there is no tactile YouTube), and a demonstration that touch-primary policies generalize across objects and tasks the way vision-language-action models are beginning to. Nobody has trained a tactile foundation policy at the scale vision has enjoyed. That gap is the opportunity, and it is exactly the gap between Ernst's dozens of bits and today's megabits per second of touch.

Where to read it

The paper is "MH-1, a computer-operated mechanical hand," Proceedings of the 1962 Spring Joint Computer Conference, DOI 10.1145/1460833.1460839 (bibliographic details verified; the fuller account is Ernst's 1961 MIT ScD thesis of the same title, available from MIT). Read it alongside Qi et al., "General In-Hand Object Rotation with Vision and Touch" / "Rotating without Seeing" for the modern vindication; Meta's Sparsh paper for tactile self-supervision; and, for the intermediate link in the chain, the 1970s IBM and Stanford work on guarded moves and manipulation languages (Will & Grossman; Bolles & Paul) — which is essentially MHI grown up, one substrate generation before it could learn.