RoboBoom
sudoremove· July 21, 2026· 33:52

Sunday Robotics ACT-2 Preview 내용 훑어보자

sudoremove discusses Sunday Robotics’ ACT-2 laundry-folding results and data strategy, then turns to Walden Robotics, founded by Russ Tedrake and former Toyota Research Institute colleagues.

In this talk

  1. 0:00

    Sunday frames laundry as a deployment problem

    Sunday Robotics says its Memo robot folds laundry with 99.1% success and describes the result as a solved task. The host says the claim depends on a defined scope: eligible garments, varied rooms and bed setups, and operation without home-specific training. A five-star rubric gives an average score of about 4.72, with blouses scoring lower than T-shirts.

  2. 4:46

    Pretraining and data quality support generalization

    The discussion links more pretraining data to a smaller performance gap between familiar and unfamiliar homes. It also highlights that carefully selected subsets can outperform uniform samples, even when training uses only a fraction of the available data. Sunday reportedly fine-tuned the same model with one example of each folding style, though the hosts question how novel those styles were to the model.

  3. 9:35

    Sunday builds its own data pipeline

    The hosts describe Sunday’s end-to-end operation: it owns the robots and models and gathers data using a wearable device called the Skill Capture Glove. Workers called Memory Developers fold garments while logging demonstrations; reported pay ranges from roughly $30 per hour on average to $60 for high-quality approved data. The hosts estimate that a workforce of 1,000 could make data collection very costly, while noting that the approach gives Sunday control over data quality.

  4. 16:58

    Memo uses edge inference for home conditions

    Sunday first tried cloud inference, but variable home Wi-Fi made it unreliable, even after experiments with Starlink. The company then built edge inference software, which the host takes as a clue that the model may be in the low billions of parameters, perhaps below 10 billion. The discussion also points to broader pretraining capabilities, including vacuum use, multi-object grasping, zipping clothes, and turning socks inside out.

  5. 25:46

    Walden Robotics focuses on workplace augmentation

    The episode introduces Walden Robotics, founded by MIT professor Russ Tedrake with colleagues from Toyota Research Institute, with Toyota reportedly supporting a proof of concept. The company plans to work in industrial settings and describes its goal as helping workers perform better rather than replacing them. The hosts cite a $300 million seed round at a $1.1 billion valuation and discuss the company’s interest in the social effects of automation.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks