Teleoperation and data factories for humanoid AI

Lead strip: Status: teleop + fleet data still core · Availability: autonomy claims ≠ unsupervised fleets · All-in cues: demo hours ≠ paid autonomy · Scoring: methodology

Evidence brief only — not a sold ranking. Teleop data ≠ autonomous field hours. A “data factory” is a training pipeline, not proof the robot already works alone on your floor.

Humanoid autonomy narratives often skip the boring middle: who moves the arms while the model learns, how many hours of which data class exist, and whether deployed robots still need remote humans. In 2025–2026 the densest public disclosures sit with Figure’s Helix / Go-Big stack, 1X’s Redwood + Expert Mode framing, and Tesla’s Optimus Academy / IR training language. This brief maps those pipelines and the buyer questions they imply. Cross-read Digit, Figure 03, Optimus.

Robot strips (context, not a trophy board)

• Digit · Agility Robotics · Status: Digit 4 field / Digit 5 announced · Use: warehouse logistics · Evidence: paid site hours/totes (ops) — training-data hours public — · → /robots/digit/
• Figure 03 · Figure AI · Status: Pilot (BMW Spartanburg sequencing) · Use: factory logistics · Evidence: Helix teleop hours + BotQ fleet-as-data · → /robots/figure-03/
• Optimus · Tesla · Status: Research / production-ramp · Use: factory R&D / Academy · Evidence: IR / Academy framing; external buy page — · → /robots/optimus/

Four data classes buyers should separate

Data class	What it is	What it does not prove
Teleoperated robot demos	Human pilots drive the humanoid; logs train imitation / VLA policies	Unsupervised shift work
Wearable / mocap / camera human demos	Human motion captured without (or with less) robot teleop	Transfer quality on your SKU
Egocentric / internet / home video	Passive video for navigation / pretraining	Dexterous contact success
Fleet deployment logs	Robots in real sites (success + failure)	That every hour was autonomous

Mixing these into one “AI trained on millions of hours” slide is how autonomy claims get oversold.

Figure: teleop → Helix → BotQ as a data engine

Figure’s February 2025 Helix paper states Helix was trained on a high-quality multi-robot, multi-operator teleoperated dataset of about ~500 hours, with a VLM generating hindsight language labels from onboard video (Helix). That is an unusually explicit supervised teleop hours disclosure — and still a small absolute number relative to internet-scale language corpora.

Later Figure posts treat hardware scale as a data strategy: BotQ-built robots are allocated to R&D, data collection, housework end-to-end work, and commercial use-case development; “the larger our fleet becomes the more data we are generating for Helix” (Ramping Figure 03 production).

Project Go-Big adds a second lane: large-scale egocentric human video in Brookfield residential and other real environments, including a claimed navigation result trained on 100% human video with no robot demonstrations (Project Go-Big). Useful for pretraining / navigation narratives — not a substitute for contact-rich manipulation hours on the robot itself.

Hiring / role language in secondary coverage separates Helix Data Creators / Humanoid Robot Pilots (training demonstrations, often with a rig) from Humanoid Robot Operators (supervise commercial sessions). Treat staffing postings as process evidence, not as published autonomy KPIs.

Buyer cue for Figure 03: BMW partner sequencing is a deployment evidence class. Helix’s ~500 teleop hours and Go-Big video are training evidence classes. Do not merge them into “BMW proves Helix is fully autonomous.”

1X: Redwood trained on teleop + autonomous episodes

1X’s Redwood post (June 2025) describes a ~160M-parameter onboard VLA for mobile manipulation, trained on a large dataset of teleoperated and autonomous episodes from EVE and NEO, including office and employee-home diversity (Redwood AI). Redwood is framed to learn from success and failure rollouts, not only clean demos.

Consumer NEO marketing and independent explainers also describe Expert Mode / remote human assistance when tasks exceed current autonomy — sessions that can feed learning loops. That disclosure is commercially honest; it also means checkout ≠ unsupervised housemaid. Score autonomy claims against whether a remote human is still in the loop for hard chores.

Further reading: House of 47, which runs Robotscomparison, looks at why today's home humanoids still lean on a remote operator.

Tesla Optimus: Academy and mixed demonstration sources

Tesla IR / earnings language in 2026 frames early Optimus builds for the Optimus Academy — internal robots collecting training data and developing functionality rather than shipping as a catalog product (Optimus brief context). Public call / update language points to a mix: factory-worker observation, dedicated demonstration teams, internet video, and reinforcement learning from robot practice — with secondary reporting that Tesla has shifted emphasis among mocap teleop, cameras, and other demo styles over time.

External catalog buy page and stable list price: —. Academy hours are not customer fleet hours.

Digit: ops proof ≠ published teleop ledger

Digit carries the densest paid warehouse operations narrative among Western programs we track (Digit 4: >65k hours, 100k+ GXO totes — vendor). Agility’s public materials emphasize Arc software, site workflows, and Digit 5 cooperative safety — not a Helix-style published teleop-hour dataset. For buyers: field ops evidence and training-data factory evidence are different columns. Digit wins on the former in public; teleop-hour disclosure remains —.

What this means for autonomy claims

1. Ask which hours — teleop, human video, sim, or unsupervised deployment.
2. Ask who intervenes — on-site operator, remote Expert Mode, or none.
3. Ask generation — Helix 01 narrative ≠ Helix 02 plant policy; Digit 4 ≠ Digit 5.
4. Ask environment — kitchen demo video ≠ BMW sequencing ≠ GXO tote transfer.
5. Refuse “fully autonomous” without a definition of supervision rate and task envelope.

A data factory can be a real competitive asset. It is still a means of training, not a certificate that the robot already replaces labor without humans in the loop.

Practical takeaway

Figure publishes the clearest teleop-hour → VLA story (~500 hours for early Helix) and is scaling fleet + human-video collection. 1X documents teleop + autonomous home-oriented training and retains remote assistance as a product reality. Tesla’s Academy is an internal data / development frame without an external buy page. Digit’s strength is paid logistics ops, not a public teleop ledger. Score each class under methodology.

FAQ

If a vendor used teleop to train the model, is the robot still “autonomous”?

Training often uses teleop; deployment may still be autonomous, semi-autonomous, or remote-assisted. Ask for the runtime supervision model, not only the training recipe.

Is 500 hours of teleop a lot?

For early Helix, Figure presents ~500 hours as comparatively small vs prior VLA corpora while still enabling upper-body control claims. Absolute adequacy depends on task diversity and evaluation — not a trophy number.

Does Project Go-Big mean Figure no longer needs robot demos?

Go-Big targets scale and diversity (including human video → navigation transfer claims). Manipulation and contact-rich skills typically still need robot-interactive data. Read Go-Big as a pretraining lane, not a full replacement for teleop / fleet logs.

Where does Digit fit?

Digit’s public edge is named-site operational hours and totes, not a published teleop-hour paper. Different evidence class from Helix’s training disclosure.

Sources

• Figure AI: Helix · Project Go-Big · Ramping Figure 03 production
• 1X Technologies: Redwood AI
• Robotscomparison: Digit · Figure 03 · Optimus · methodology

Exact remote-intervention rates, live BMW autonomy %, and Tesla Academy hour tallies: — where not public.