DYNA Robotics' Towel-Folding Robot Swaps Task Success Rate for MTBI as the Key Metric

Forum topic · QianXun · 2026-10-01

Summary

On September 29, DYNA Robotics unveiled DYNA 2.1, a physical agent combining a dual-arm upper body with swappable grippers and a four-wheel steering base, targeting full commercial laundry shifts: loading washers, extracting towels from dryers, shaking, folding, sorting by size, and shelving autonomously. The post's central argument is mathematical: with 1,000 sequential steps per shift, a 99% single-step success rate yields only ~0.004% fully correct shifts, while 99.9% yields ~36.8%—meaning single-episode success rate is the wrong unit for measuring real-world viability. DYNA therefore adopts MTBI (Mean Time Between Interventions) as its optimization target, logging decision traces down to joint loads and feeding them back into training. Reported deployments include hotels, laundromats, and restaurants, with existing Dintaifung (Din Tai Fung) restaurant deployments claimed to have achieved ROI and expansion to hundreds of units planned by H1 2027. However, DYNA has not disclosed its current MTBI reading or statistical methodology, and averages can mask intervention distribution patterns. The company was founded by Lindon Gao and York Yang (who sold Caper AI for $350 million) and Jason Ma (ex-DeepMind), backed by CRV and First Round.

A quick math problem first

Assume a shift contains 1,000 consecutive steps, each independently successful:

Single-step success 99% → probability all 1,000 steps succeed ≈ 0.99¹⁰⁰⁰ ≈ 4.3×10⁻⁵ (about 4 in 100,000)
Single-step success 99.9% → ≈ 36.8%
Single-step success 90% → ≈ 1.7×10⁻⁴⁶

This requires no paper citations. The point: describing a full shift with single-task success rate is a unit mismatch.

The robot released on September 29

DYNA Robotics released the DYNA 2.1 physical agent on September 29 via PR Newswire; RoboticsTomorrow ran a preview the day before without model name or specs.

Hardware: an upper body with two arms, grippers swappable between parallel-jaw and dexterous hands, on a base with four steering wheels. Wheels sidestep all bipedal balance problems—at the cost of unreachable heights and no stairs.

Software, three layers: 1. A vision-language orchestrator reasoning over multi-step workflows 2. A whole-body controller merging chassis movement, arm extension, and bending/lifting into single continuous motions 3. The DYNA-2 world action model underneath, trained on one million hours of human and robot data

What MTBI replaces

Co-founder Jason Ma's argument: isolated single-episode task success rates don't reflect real commercial viability, since real shifts demand thousands of consecutive steps. CEO Lindon Gao: customers want a full-shift robot that completes entire workflows without a human standing by.

The company switched to MTBI (Mean Time Between Interventions) as the optimization target, logging data from orchestration-level decision traces down to control-level joint loads and feeding it back into the system.

A complete laundry-shift pipeline

The publicized scenario is an entire commercial laundry shift: load and start the washing machine, pull towels from the deepest recesses of the dryer, shake flat, fold, sort into piles by size, and shelve on racks of different heights. Towels dropped on the floor get picked up autonomously; failed grasps trigger a release-and-retry—all without asking a human.

Verifiable deployment facts

Deployed settings: hotels, laundromats, and restaurants. Both the press release and follow-up coverage cite one piece of commercial evidence: existing deployments at the Din Tai Fung restaurant chain have reportedly achieved ROI, with expansion to hundreds of units planned by H1 2027.

| Item | Disclosed by DYNA? | |---|---| | Training data scale | 1M hours, human + robot combined | | Hardware form | Dual arms + four steering wheels, disclosed | | Deployment settings | Hotels, laundromats, restaurants, disclosed | | Commercial ROI evidence | Din Tai Fung ROI; hundreds of units by H1 2027 | | Current MTBI reading | Not disclosed | | MTBI statistical methodology | Not disclosed; what counts as an intervention is undefined |

Founding team: Lindon Gao and York Yang, who previously sold Caper AI for $350 million; Jason Ma, formerly a DeepMind research scientist. Investors include CRV and First Round.

Why changing the metric matters more than the hardware

MTBI's advantage maps directly to cost: one human intervention costs one person's time, so MTBI times labor rates gives the variable portion of shift cost—something a financial model can ingest.

Its gap is equally clear. Averages hide distributions: one intervention per hour across ten hours versus ten interventions crammed into the first hour can share the same mean but mean entirely different things for scheduling. The definition of an intervention is also unbounded—does a hand touching the robot count, or does a single remote screen tap?

The company made MTBI its optimization target but hasn't published its own current reading on that target. At minimum, outsiders cannot judge whether DYNA itself meets its own bar. The metric is its invention—and the first report card is one it gets to keep.

---

Sources: PR Newswire official release (Redwood City, 2026-09-29), The AI Insider (2026-09-30), RobotToday, FairsOnline, and a RoboticsTomorrow preview from September 28.

Caveats: The 1M-hour training data figure and MTBI improvements are company self-reported; Din Tai Fung figures were not verified with the company itself; no current MTBI value or statistical methodology was given in the release; the opening probability arithmetic is computed independently under i.i.d. assumptions, not from a paper.