Electronically controlled manipulator

TL;DR: In 1954, Ray Goertz and William Thompson at Argonne built the first electrically servo-coupled master–slave manipulator with force reflection: an operator moved a master arm, a remote slave arm mirrored it, and — crucially — forces on the slave were reflected back so the operator could feel the hot cell through the wall. The idea was decades ahead of its actuators, its electronics, and above all its economics: one human per arm, forever, with nothing learned and nothing kept. Today teleoperation logs are the training corpus for manipulation foundation models, and Goertz's central claim — that force coupling is what makes remote dexterity possible — is exactly the signal most modern data pipelines throw away.

The idea as Goertz had it

The problem was radiochemistry. After the Manhattan Project, national labs needed to manipulate intensely radioactive material behind meters of concrete and leaded glass. Goertz's first solution (late 1940s) was purely mechanical: master and slave arms joined by cables, tapes, and rotating shafts through the wall. These worked remarkably well precisely because they were mechanically transparent — push on the master, the slave pushes; the slave hits a wall, you feel it. But mechanical linkage caps the distance at a few meters, constrains the geometry to "operator directly outside the cell," and leaks radiation through every penetration.

The 1954 paper's move was to cut the mechanical link and replace it with wires: position sensors (synchros) on both arms, servo amplifiers, and motors on both the master and the slave. Each joint runs what we'd now call symmetric position–position bilateral control. Let x_m and x_s be master and slave joint positions. Both motors are driven by the same error:

When the slave moves freely, the error is near zero and the operator feels almost nothing. When the slave contacts something, x_s lags x_m, and the growing error pushes back on the operator's hand with the same stiffness K it pushes the slave forward. Force reflection falls out of the symmetry — no force sensor required. The environment becomes, in effect, a spring the operator can feel from the other side of a concrete wall, or in principle from another building entirely.

Against the state of the art in 1954, this was radical. Industrial "robots" did not exist (Unimate was seven years away). Feedback control theory was fresh — servomechanism theory had been systematized during WWII for gun turrets and radar. Goertz was, in effect, proposing the internet-era architecture of remote manipulation — sensing, transmission, actuation, feedback in both directions — while working with synchros and vacuum tubes.

Why it couldn't work then

Bandwidth and fidelity. The analog servo chain — synchro pickoffs, tube amplifiers, wound-field or early servo motors driving high-ratio gear trains — gave a useful closed-loop bandwidth of roughly a few Hz. Human contact perception is much faster: kinesthetic sensing to roughly 20–30 Hz, vibrotactile sensing up to roughly 1 kHz. Reflected friction and inertia from gearboxes swamped delicate contact forces. Operators of the electric arms consistently reported them as inferior in "feel" to Goertz's own mechanical linkages — the electric version bought distance at the price of transparency. Goertz spent the rest of his career (he died in 1970) trying to claw that fidelity back, and largely could not with the components of his era.

Stability. Position–position bilateral control with stiff coupling and any transmission delay is a textbook instability. The theory to handle delayed bilateral teleoperation — passivity arguments, scattering/wave variables — arrived in the late 1980s (Anderson & Spong 1989; Niemeyer & Slotine 1991). Goertz had to keep gains low and arms local, which undercut the whole promise of arbitrary-distance operation.

No memory, no leverage. This is the deep one. Every demonstrated behavior evaporated the moment the operator's hands left the master. There was no practical way to record joint trajectories and force profiles (instrumentation tape existed, but replaying a trajectory open-loop in a contact task is useless), and no conceptual framework for turning demonstrations into controllers — "programming by demonstration" is a 1980s–90s idea, and learning general policies from demonstration logs is a 2020s reality. So the economics were fixed: one skilled human, one arm, one hour of attention per hour of work. Teleoperation could never amortize.

Cost. A hot-cell servomanipulator installation was lab-scale capital equipment — plausibly the equivalent of hundreds of thousands of today's dollars per station. Fine for plutonium, absurd for anything else.

What changed

Every limiting component flipped. Brushless motors with rare-earth magnets and quasi-direct-drive transmissions give high torque density with low reflected inertia — you can build a master arm that feels nearly transparent. Cheap encoders and microcontrollers run torque loops at 1–2 kHz, well above kinesthetic bandwidth. A capable bimanual teleop rig (ALOHA-class: two puppet masters, two follower arms, cameras) costs on the order of $20–30k, and stripped-down leader-arm rigs like GELLO cost a few hundred dollars — roughly three to four orders of magnitude cheaper per station, in real terms, than Goertz's installations.

But the decisive change is what happens to the logs. Imitation learning — ACT, diffusion policies, and now vision-language-action models like π0 and OpenVLA — turns teleoperation from a consumption of human attention into an investment. One hour of demonstration becomes training data for a policy that runs unattended thereafter. The Open X-Embodiment and DROID datasets, and every frontier lab's internal manipulation corpus, are overwhelmingly built on teleoperation. Goertz's economic dead end became a data engine.

Here is the irony worth dwelling on: most of that modern pipeline is unilateral. ALOHA's leader arms give incidental kinesthetic feedback through their own joint friction; GELLO gives essentially none; VR-controller teleop gives none. The industry rebuilt Goertz's 1948 architecture minus his 1954 insight. And Goertz's own experiments (and decades of teleoperation literature after him) found force reflection markedly improves speed and success in contact-rich tasks — my recollection is that reported speedups were around 2× for tight-tolerance tasks, though I wouldn't defend a precise figure.

What a serious 2026 revival looks like

Take the premise at face value: force reflection is a first-class learning signal, and fleet-scale bilateral teleoperation is the data-collection layer for generalist manipulation.

• Hardware: bimanual quasi-direct-drive follower arms with joint torque sensing (or good current-based torque estimation), paired with motorized leader arms — Goertz's symmetric architecture, or a 4-channel scheme exchanging both position and force in each direction. 4-channel bilateral control (Lawrence 1993; refined extensively by Ohnishi's group in Japan) achieves near-ideal transparency and is the correct modern replacement for Goertz's position–position scheme.
• What to reuse from the paper: the symmetry principle (both arms actuated, environment forces felt without wrist force sensors), and the framing of the operator as in the loop of contact dynamics, not just issuing pose commands.
• What to replace: analog servos with kHz digital torque control; local wiring with passivity-based or model-mediated control over real networks; and above all, the "operate forever" endpoint with a logging-and-learning endpoint.
• The learning bet: log (x_m, x_s, \tau_m, \tau_s) at kHz alongside vision, and train policies whose action space includes force/impedance targets, not just positions. The hypothesis: operators who feel contact demonstrate qualitatively better contact strategies (regulated insertion forces, compliant sliding, controlled scraping), and force-conditioned policies inherit them — where position-only clones fail on anything with tight tolerances or delicate parts.
• The experiment: same fleet, same tasks (connector insertion, food handling, fabric, assembly), demonstrations collected with and without force reflection, policies trained with and without force in state and action. Measure success rate, applied-force distributions, and generalization under object variation. This is a clean, decidable study, and to my knowledge no one has run it at fleet scale.

Has it been tried? What's still open?

Goertz is thoroughly vindicated as the founder of a field: hot-cell servomanipulators, undersea and space teleoperation, and the entire discipline of haptics descend from this paper. The most commercially successful descendant, the da Vinci surgical system, pointedly omitted force feedback for decades — evidence that vision plus human adaptation covers a lot, but surgeons have complained about it for just as long, and force feedback appeared in newer generations.

On the learning side, the closest direct descendants are the "bilateral control-based imitation learning" line from Japanese groups (Sasagawa, Adachi, and colleagues, building on Ohnishi's 4-channel work), which trains policies on synchronized master/slave position and torque signals and reports strong results on contact-rich tasks — fast wiping, drawing, food manipulation. It is compelling but small-scale. Meanwhile, the large-scale efforts (Open X-Embodiment, DROID, π0's data engine) are essentially force-blind. The open questions:

1. Does force reflection during collection measurably improve learned policy quality at scale, or only teleoperator experience? (Small studies suggest yes to both; nothing definitive.)
2. Can consumer-grade force feedback (kinesthetic-lite, vibrotactile substitution) capture most of the benefit, or does the data engine need true 4-channel rigs?
3. How should force enter the policy — as observation, as action (impedance targets), or both — and does it transfer across embodiments?

Where to read it

The paper is Goertz & Thompson (1954), "Electronically controlled manipulator," available via OSTI: osti.gov/scitech/biblio/4387399 (bibliographic details verified in OpenAlex; it's a short trade-journal-style report, in Nucleonics if memory serves — I'm less certain of the venue than of the content). Read alongside: Hokayem & Spong, "Bilateral teleoperation: an historical survey" (Automatica, 2006) for the sixty-year arc; Lawrence (1993) on transparency and the 4-channel architecture; Zhao et al., "Learning fine-grained bimanual manipulation with low-cost hardware" (ALOHA, 2023) for the modern data engine; and Sasagawa et al. on bilateral-control-based imitation learning for the most direct attempt to close Goertz's loop with a neural network where the human used to be.