---
id: "lib-2026-10-19-dyna-robotics-revisited-after-300-days-of"
type: "library"
title: "Dyna Robotics revisited after 300 days of deployment lessons"
title_original: "再访Dyna Robotics：被部署教训了300天后，找到机器人Scaling Law"
lang: "ZH"
org: "硅谷101 (Dyna Robotics)"
source_type: "Podcast"
date: "2026-10-19"
week: "2026-W43"
url: "https://nosov.ai/library/lib-2026-10-19-dyna-robotics-revisited-after-300-days-of/"
original_url: "https://www.youtube.com/watch?v=pC3QbOu7_s4"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "A long, technically literate interview with Dyna Robotics co-founder Jason Ma that gives rare deployment numbers from a paying customer; all figures are founder claims, and they conflict with another founder's figures from the previous week."
summary_basis: "transcript"
original_minutes: "120"
tags: ["robotics deployment", "workflow vs task", "scaling", "customer requirements"]
---

# Dyna Robotics revisited after 300 days of deployment lessons

Original title: 再访Dyna Robotics：被部署教训了300天后，找到机器人Scaling Law

Tier B (Useful, limited): A long, technically literate interview with Dyna Robotics co-founder Jason Ma that gives rare deployment numbers from a paying customer; all figures are founder claims, and they conflict with another founder's figures from the previous week.

Summary based on the full transcript.

## In short

A two-hour Mandarin video podcast from Silicon Valley 101 (19 October 2026) with Jason Ma, co-founder and chief scientist of Dyna Robotics, which builds robots for napkin folding and laundry. Dyna treats real deployment as "the only evaluation". Its first model reached 99.4% success in the lab, but at a restaurant customer only 75% of folds met the customer's grade, and throughput was a third of what the customer needed. Dyna-2 replaced the vision-language-action model with a world-action model pre-trained on up to a million hours of human video, and met the throughput target.

## Key numbers

- 99.4% vs 75%: Dyna-1 lab success rate versus share of folded napkins that met the customer's grade
- 35/hour, ~480/day: Dyna-1 napkin throughput, against a customer requirement of 1,500 a day over an 18-hour shift
- 95/hour, 1,590/day: Dyna-2 throughput on the same task, meeting the requirement
- 100 h to 1M h: range of human first-person video used in pre-training; each 10x step improved prediction of unseen robot actions

## What it reports

- Deployment pilot: the model ran for 5 minutes, then 1, 8 and 24 hours over about 5-6 weeks, learning from its own rollouts and targeted error-recovery data.
- The customer grades folds 1-5 and accepts only 4-5; the gap between grades is millimetres. Dyna-1 hit 99.4% lab success, 75% at customer grade.
- At 1,000 repetitions, 1% failures still need human or machine QA downstream, so a 99% robot adds work. The aim is a robot that owns a whole workflow without QA, even if slow.
- Throughput: Dyna-1 35 napkins an hour, about 480 a day; requirement 1,500 a day over an 18-hour shift; Dyna-2 95 an hour, 1,590 a day. Going from 1,400 to 1,500 had "no low-hanging fruit".
- Scaling: pre-training on 100 h, 1k, 10k, 100k and 1M hours of human video, each 10x improved prediction of robot actions unseen in pre-training; gains flattened from 100k to 1M for lack of high-quality labelled video. Most video was bought from partners at a cost "far below teleoperation".
- Post-training for commercial reliability still needs teleoperation data; Dyna-2 learns a new task from tens of minutes of data.
- Intervention paradox: a faster stationary robot needs staff to clear its table more often, which pushed Dyna toward a robot that runs the whole laundry flow.
- Strategy: hardware plus deployment is the moat; models are swappable. B2B sites before homes. Team of 70-80 people in the US.

## How they got it

An interview. The guest describes internal experiments and a customer pilot; no papers, curves or customer confirmation are shown in the transcript. The digest is based on the uploader's Chinese captions with English glosses.

## Limits

- All figures are the founder's and unverified. Last week's interview with another Dyna co-founder gave different throughput figures for the same pilot; the conflict is unresolved.
- The stated 35 an hour and about 480 a day do not reconcile with an 18-hour shift; the shift length may differ by phase.
- The scaling claim is described verbally; data mix and evaluation set are not disclosed.

## Why it matters for companies

The gap between 99.4% in the lab and 75% at the customer's quality bar is a clean illustration of why pilots must be judged on the buyer's acceptance criteria and throughput, not vendor success rates. The point that 1% failure at volume still needs a QA step applies to any automation, physical or software. Buyers of robotics should ask for grade-level acceptance and shift-level throughput figures from their own site before committing.

## Key passages in translation

> 实验室的成功率是做到了99.4%但是只有75%的叠餐巾是达到了后来真实的客户标准

The lab success rate reached 99.4%, but only 75% of folded napkins met the real customer standard.

> 最开始的时候Dyna-1是每个小时叠35个纸巾大概是每天可以叠480个但是当时客户就对你们的要求是说希望你们可以做到每天1500个

At first Dyna-1 folded 35 napkins an hour, about 480 a day, but the customer's requirement was 1,500 a day.

> 真实部署是衡量机器人的“唯一评估”

Real deployment is the "only evaluation" of a robot.

Original: https://www.youtube.com/watch?v=pC3QbOu7_s4