What you would work on
We define three levels of ownership on the ML team. Which one you start at is decided through the interview process.
You are given a specific technical theme and improve performance through it
e.g. applying a promising paper, or data-side improvements such as camera tuning
You are given a partial goal and reach it — selecting existing techniques and, where needed, proposing new ones
e.g. reducing the failure rate of simple towel grasping during autonomous towel folding
Given the goal (raising the revenue a robot generates per hour), you propose what the ML team should do in the first place
e.g. proposing a data collection strategy, or which methods we should be trying
We measure results in profit generated per robot per hour. The experiments we actually run are published openly, including the ones that did not work.
Read our research logRequirements
- 01
Able to come to our factory in Shikahama, Adachi-ku, Tokyo at least twice a week
We operate four bimanual robots at our own factory. On-site presence is for attending physical experiments and for in-person discussion. In robotics, improvement ideas often come from seeing the actual environment and behaviour.
- 02
At least one year of professional machine learning experience
Research conducted as part of a degree does not count.
- 03
Mathematical grounding sufficient to deeply understand the architecture and training methods of VLA models such as pi0.5
This is needed so that when a newly implemented method does not deliver the expected result, you can diagnose the cause accurately and improve on that basis.
- A closed-book written exam during the interview, covering derivations in linear algebra, probability and optimisation
- Oral questions in the interview — e.g. the training objective of the flow matching action head in pi0.5, or what knowledge insulation is preventing
- 04
Able to hold technical discussions in English (spoken)
Reading papers and writing — Slack posts, reports and so on — assume the use of AI, so we only require spoken communication.
- Japanese language ability is not required.
Nice to have
- Experience with deep learning (especially VLA) and robotics
- Experience with teleoperation and data collection on physical robots
- Experience operating training infrastructure (multi-GPU training, cloud GPU cost management, data pipelines)
- A research publication record
Terms
- Employment type
- Full-time
- Annual salary
- JPY 8,000,000 – 15,000,000+
- Location
- Shikahama, Adachi-ku, Tokyo (our own factory)
- On-site
- At least twice a week
- Team
- 2 ML engineers, 4 data engineers
- Compute
- 10+ in-house GPUs, plus cloud GPUs
- Hardware
- 4 bimanual robots
Hiring process
Application review
Send us your CV.
Casual conversation (once)
For mutual understanding. You are not being evaluated.
Interview (once, 120 minutes)
60 minutes of technical discussion, plus a 60-minute written exam.
The written exam is closed-book: no smartphones, no PCs, no LLMs. Its scope and difficulty match the sample problems below.
Sample assignment ver 1.0
These are not the exact problems you will be given, but the scope and difficulty are the same. No submission is required. We publish it in full so that you can see what we are measuring before you apply.
- Problem 1
Linear algebra on the self-attention score matrix, and derivation of its gradients
- Problem 2
Conditional expectation of a linear interpolation between two distributions (the flow matching velocity field)
- Problem 3
Interpreting real internal experimental data and proposing a research direction from it
Problem 3 uses data from an experiment we actually ran on the RPE (Remove Pillow Envelope) task. The success rates did not come out the way we expected, and we publish them as they are.
Download the sample assignment ver 1.0 (PDF, 1.5MB)JOIN US
Talk to us first
It is fine if you are unsure whether you meet the requirements. The casual conversation is not an evaluation.