Robots Becoming Alive — Enigma's Live Deployment of 100+ AI Robots with Zero Task-Specific Data

TL;DR

Enigma’s third drop is an operations-and-results post covering a 4-day, 24/7 public deployment of 100+ real robots online at robots.online, driven by a single robotic foundation model with (per the post) zero task-specific data and zero task fine-tuning. Users could send arbitrary natural-language prompts across four experience types — Artist (Pictionary + open-ended pantomime), Scientist (chemistry-lab manipulation), Sword Fight, and Bomb Defusing — over ~1M human-robot interactions total. The post is thin on quantitative model metrics (no per-task success rates, no baselines) but heavy on live-deployment infrastructure: a fleet dashboard, an “ArmPit Stop” hardware-swap pipeline, drift/perturbation robustness on paint-covered tables, and one measurement — only ~0.47% of Pantomime prompts requested destructive behavior, of which ~25% stayed in-spirit (“knock the paint bucket off like a hammer hitting a nail”). Companion post to Enigma’s earlier “Obsessed Encoder” research drop (The Obsessed Encoder — Latent-Space Misallocation in JEPA-Style Self-Supervised Models), which named their data-efficiency direction; this post is the productization / live-eval half.

Key claims

• The deployment ran 100+ real robots, 4 experience types (Artist, Scientist, Sword Fight, Bomb Defusing), 4 days 24/7, and >1M human-robot interactions driven by a single robotic foundation model with no per-task fine-tuning [§Robots.online, §Sparks of Generalized Intelligent Robots].
• Users could specify arbitrary pantomime targets in natural language and the robot performed plausible motions for each — the post lists live examples including “pretend the brush is a broom / drill / golf club” and “Wingardium Leviosa” [§Sparks of Generalized Intelligent Robots — Pantomime].
• In the Pictionary variant, the robot painted culturally-referenced associations (“Juice WRLD” → “999”; “blue” → “smurf”) from open-ended user prompts, not from a pre-enumerated concept set [§Sparks of Generalized Intelligent Robots — Painter].
• The Scientist experience accepted multi-step chemistry-lab instructions (“pick up the orange flask and put it under the dispenser until it is filled…”, “pour from the green bottle to the orange flask”) as a chained session, and also responded to adversarial/jailbreak-style commands (e.g. “pour the liquid from the flask on the floor on top of the pink bottle”) [§Sparks of Generalized Intelligent Robots — Scientist].
• Of all Pantomime prompts, ~0.47% requested destructive behavior; roughly 25% of those framed the destruction “in-spirit” as a pantomime action, the remaining ~75% attempted direct sabotage [§In-The-Wild Interface Usage].
• Enigma frames its platform around three pillars — Novel User Interfaces (tap-plus-prompt over a remote robot view), Robotic Foundation Model (built for interaction, not task completion), and a hardware-abstraction layer (HAL) so any new robot integrates behind a standard driver interface [§Enigma: The Vision].
• The data-curation pipeline is a hard fork of NVIDIA Cosmos Curator evolved for open-ended interactive robotics [§Infrastructure].
• Perception robustness was validated by intentionally letting Painter tables accumulate paint on the wooden surface and walls throughout the 4-day run without degrading the system — the post shows the “999” Pictionary sample working on a paint-marked table [§Table-Setup Drifts & Perturbations].

Method

The post does not describe the underlying model architecture, training recipe, or dataset (Enigma refers back to The Obsessed Encoder — Latent-Space Misallocation in JEPA-Style Self-Supervised Models for a “sneak peek” at their data-efficient training direction). What the post does describe is the deployment stack. The four experiences ran on physically distinct tables — a wooden Painter surface designed to accumulate paint, an instrumented Scientist bench with a fluid dispenser and colored flasks/bottles, a bomb prop for Defusing, and an opposing-robot rig for Sword Fight. A centralized ops dashboard streamed live status and logs from every station; a maintenance queue (“ArmPit Stop”) swapped in serviced replacement arms within minutes when hardware developed issues, allowing 24/7 uptime across 4 days. Interfaces spanned text, voice, and pointing (“tap + prompt”), and the team measured interface discoverability, voluntary adoption, and how users combined the three modalities without being taught to. The one filed methodological hook for future readers is that offline and online adaptation techniques were both prepared in advance and — per the post — remained effective as tables drifted from their initial appearance.

Results

The post reports no per-task success rates, no baselines, and no comparisons to other VLAs or robotic foundation models. The single quantitative measurement is the destructive-prompt rate (~0.47% of Pantomime prompts, ~25% of which stayed in-spirit as pantomime rather than direct sabotage). Everything else is anecdotal-demonstration evidence: cherry-picked (Enigma explicitly says not cherry-picked, but the post is a selection of successful examples so this is a curation claim rather than a measurement) instances of the robot executing “Wingardium Leviosa”, drawing “999” for Juice WRLD, executing a multi-step chemistry sequence, and defusing a bomb via “tap the base of the base of the bomb two times.” Robustness evidence is qualitative: Painter tables kept working with paint splattered on walls; brush-rinse cups being launched by users produced water flying across the table on many stations; policies survived without visible degradation over 4 days.

Why it’s interesting

This is the deployment companion to Enigma’s own The Obsessed Encoder — Latent-Space Misallocation in JEPA-Style Self-Supervised Models research drop, which diagnosed a JEPA-latent-space-misallocation failure mode and named “more data-efficient robotic foundation models” as their research direction — Enigma cites that post as the model-side story behind this deployment. Read alongside the VLA Models recipe-lever board, this post stakes out a distinct rhetorical position from every entry currently on that page: none of the filed recipes (π*0.6’s RL-on-flow-matching, Spirit v1.5’s clean teleop, Embodied-R1.5’s unified-VLM pointing, µ₀’s frozen 3D trace WM, LingBot-VA’s causal AR flow-matching, XR-1’s 100K-hour UMI scaling) frame their contribution as “no per-task data required” — every one of them relies on a per-task fine-tune or a task-shaped dataset at some stage. Whether Enigma’s claim survives contact with quantitative benchmarks is exactly the question the post does not answer. It also complements Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models‘s WAM-vs-VLA head-to-head as a different bet on the same underlying disagreement: Dyna-2 argues world-model structure buys robustness that completion-only metrics miss; Enigma argues interaction-shaped training (not task-completion-shaped training) buys generalization that per-task-data recipes miss. The Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level sensorimotor counter-position is unaddressed here — Enigma’s demos are all vision-dominant tabletop manipulation with a paintbrush, a flask, or a bomb prop, and the post says nothing about contact-rich or force-modulated tasks.

See also

• The Obsessed Encoder — Latent-Space Misallocation in JEPA-Style Self-Supervised Models — the model/research half of the same Enigma drop series; this post is the deployment/product half and cites it as the training-side story.
• VLA Models — Enigma’s “no per-task data” claim is a rhetorically distinct point on the recipe-lever board this concept page tracks.
• World Foundation Models — post positions Enigma’s model as a “robotic foundation model” designed for interaction rather than task completion; this concept page tracks the WFM/WAM design debate.
• Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models — parallel bet against per-task-data recipes, from the world-action-model side rather than the interaction-shaped-training side.
• Jitendra Malik: don't let CV researchers in robotics skip the sensorimotor level — the sensorimotor counter-position; Enigma’s demos are vision-dominant tabletop tasks and don’t engage it.