okumasan

IT企業でロボットのR&D

[よこしゃん｜ヒューマノイド最前線🤖 @yokoshaan]

両腕ロボットが中国の七弦琴「古琴」で25音のフレーズを自律演奏し、弦を押さえて響かせる奏法で音の正確さ96.8%を出しました。

古琴はフレットがなく弦の間隔も数ミリで、右手ではじく一瞬のタイミングと、左手で押さえ続ける位置の両方が音を決めます。

カメ���で楽器の位置を合わせ、触覚で押さえ具合を見張り、実際に鳴った音を聞いてはじく強さを調整します。開放弦だけの列では93.6%でした。

[Skild AI @SkildAI]

Introducing S1, our new foundation model that learns from one example.

It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning.

Watch S1 operate in real-time via in-context learning:

[NVIDIA Robotics @NVIDIARobotics]

140 years of soccer practice in simulation. ⚽

Skild AI’s robot policy learned to dribble, shield, and tackle through self-play in NVIDIA Isaac Sim before being transferred to a real humanoid robot. 👇

[YUJI | AI Robotics @yuji_fujima]

【フィジカルAIでサッカー】

米Skild AIから新しいデモ

同社のロボット基盤モデル「S1」をベースに、 対戦を繰り返して学習

与えた目的は「ゴールを決める」だけ

・ドリブル
・ボールキープ
・タックル
を対戦の中で自然と獲得

フィジカルAI、どんどん進化してる

[GREG ISENBERG @gregisenberg]

I think we'll see 5x the number of people become hardware and robotics founders over the next 18 months

Why? Because you can rent every step, from design to manufacturing:

DESIGN

- Astra to drive Blender, FreeCAD, and KiCad the way a person would, so you get editable geometry instead of a dead render. Zoo dot dev's text to CAD API if you want it programmatic.

- Or have Claude write build123d code, which is Python parametric CAD, and there's a Claude Code plugin that runs it and exports the STL for you.

PROTOTYPE

- Bambu A1 mini is $299, P1S around $399 on sale. Or skip owning one, JLC3DP prints and mails it for about $20. PCBWay does resin, SLS, and CNC if plastic won't cut it.

ELECTRONICS

- KiCad for the board, Astra can drive it. JLCPCB fabricates and assembles, often under $100 for a small run. ESP32 for wifi, Raspberry Pi if it needs a brain.

ANYTHING THAT MOVES

- Unitree sells a Go2 with a full SDK for about $2,500.

- Hugging Face put out a $399 open-source biped.

- LeRobot gives you the whole train in sim, deploy to real pipeline, and Physical Intelligence open-sourced π0 so you're fine-tuning instead of starting from zero.

- Prototype the policy in MuJoCo or Isaac Lab first.

MANUFACTURING

- Alibaba RFQ for the first 100 units. Check 1688 to see what the factory actually charges domestically, then negotiate. Pietra or Sourcify if you want someone to handle it.

FULFILLMENT AND SELLING

- ShipBob or Amazon FBA. Shopify for the store, TikTok Shop for distribution, Kickstarter if you want the money before you build it.

How to think about starting your own robotics or hardware company:

1. Pick a niche that's already buying weird gear. Cyclists, tabletop gamers, beekeepers, home baristas, dog people with mobility issues. These groups spend money on specific objects and complain in public about what doesn't exist. Or grab ideas off https://t.co/QxoITW16zr

2. Go read the complaints. Reddit, IG etc Search "I wish someone made," "does anyone make," and "modified my." That last one is the best signal, because someone already hacked the product together and you're just manufacturing what they built by hand.

Also check Etsy!! If 3 sellers are doing a janky 3D printed version with 400 reviews each, the market is validated.

3. Make one. Describe it to Astra or Claude, get the CAD, print it in ugly gray PLA, use it, fix it.

4. Only go to Alibaba or similar once you've sold a few. Message 10 suppliers through RFQ, take the third cheapest, always pay the $50 for a sample before the real order.

5. Film everything from day 1. The first ugly print, the failed version, the box of 100 arriving. That's your entire marketing budget, and hardware is one of the few categories where people actually want to watch the thing get made.

6. Raise the price. Almost everyone here anchors on what the plastic cost. Your customer is comparing you to nothing, because the alternative is the product doesn't exist. Start at 5x COGS and go up.

THIS GOAL OF THIS POST IS JUST HERE TO GET YOUR CREATIVE JUICES FLOWING. Of course, you can build robotics/hardware in a bunch of different ways.

One thing I've learned is data couldn't be more important when you're building a hardware/robotics startups.

These models learn from first person video of a human doing the task, and that footage doesn't exist for almost any job.

So basically you pick one repetitive job people quit over, film someone doing it for 2 weeks, fine-tune π0 on that footage, use organic to sell the first few units and figure out scaling,

5 years ago you needed a factory, a supply chain, and a $1M just to find out if anyone wanted the thing.

Now, anyone can become a hardware/robitics founder.

And I suspect a lot of people will become one!

Vibe manufacturers.

[Yoshihiro Shibata @shiba_8ro]

TOYOTA&AIRoAでフィジカルAI向けデータ収集グローブ&ロボットグリッパーな「YUBI」がシルバーウィーク中にV2に進化したわけですが（トヨタカレンダーでの稼働お疲れ様です！）V1との差分をみるとデータ収集グローブのフィンガー部品がギヤと一体型になったのがメカ的にはハイライトな模様👀一方でRSJのTOYOTAブースで見かけた液晶パネル搭載版グローブな改造はマージされなかった模様 ＃YUBI

[よこしゃん｜ヒューマノイド最前線🤖 @yokoshaan]

延世大学とRLWRLDが、人の動画をロボット視点に作り変えた63万件で学習し、実機の達成率を51.5%から80.3%に上げました。

実機でデータを集めるのは高くつき、人の作業動画は豊富ですが体の作りが違います。この研究は人の動画をロボットから見た映像と動作の軌道に変換し、欠けている中間の情報も推定して補います。

5つの動画ソースから作った1.42億フレームで事前学習し、双腕ロボットの実機4タスクで検証しました。初めての環境でも達成率が34.9%から72.2%に上がっています。

[Dan Greenheck @dangreenheck]

Ever wonder what $1874.40 of Opus 5.5 tokens looks like?

Wonder no more.

I highly recommend watching the whole video on 1080p on a large screen with sound on.

Pay attention to all the little details.
- Birds flying and diving into the ocean.
- Cloth, signs and lights swaying in the wind.
- Crabs scuttling along the shore and burrowing when you get close.
- Fish swimming alongside the whale.
- Lights illuminating the dock at night

Then zoom out and see the entire island. Never dipping below 60 FPS at 1440p resolution.

Yes, there's a few bugs and visual artifacts, some textures need improving, the shorelines waves sometimes look funny, but these are so trivial to fix at this point.

FYI I'm on the $200 subscription plan. This used 59% of my weekly usage and took 2h 7h of API time using multiple subagents, about 8h in real-time.

There was extremely little technical direction here. 99% of my prompts were "Add X and Y" or "This looks weird, make it better".

It's a great time for hobbyists, bad time for professionals. This experiment has further cemented my view that technical creatives are about to experience a massive disruption.

[Meng To @MengTo]

Opus 5.5 is next level. I’m genuinely impressed by the three.js details.

I generated a playable boat scene through Japanese landscapes with dynamic weather, day and night lighting, realistic textures and 3D characters.

Live site: https://t.co/YMW4Zp52In

The water reflections, physics, scenery and architecture are insane. I could look at this all day.

[松本良太 (Ryota Matsumoto) @ryma_jp]

このマングローブのカヤック体験サイトは、GPT-6 Astra/Solで作りました。

自動航行は約1分、手動で漕ぐこともできます。結構リアルに再現できていると思います。Opus 5.5も凄いですが、GPT-6も戦えますね。

リプ欄のURLから体験できます↓

[Skild AI @SkildAI]

We trained a robot to play football.

How? Self-play for 140 years in a virtual World Cup.

Meet the #Messinator:

[ハヤシモン｜AI × 個人開発 @hayashimon1]

Opus5.5とThree.js
清流の表現がすごい。
透明な水、川底の光、苔むした岩まで全部コード。

難しすぎて避けてた表現が今なら全部いける気がしてる。

[SB Intuitions @sbintuitions]

RoboticsチームのVLAが、新たに「箱の組み立て」タスクを完全自律で実現しました。📦🚀

Our Robotics team's VLA has successfully achieved fully autonomous box assembly. 📦🚀

#SBIntuitions

[太田賢一／Design Mgr @kenichiota0711]

これは素晴らしいskillですね...！実際使ってみたら、自然な日本語文章が書けました。
https://t.co/saMmH0SySZ

[はさまるスタジオ (Hasamaru Studio) @hasamaru_studio]

GPT-6 AstraでBlenderを操作し、プリビズからAI動画を作る流れをよく見かけるようになりましたね！

でも、風景の世界観を再現するには、3Dをどこまで作り込めばいいのか？

同じ街並みを3段階の詳細度で作り、元画像の質感や雰囲気がどこまで反映されるか比較しました！

詳細は引用の記事にまとめています👇
https://t.co/4JXbZ9wIsY

[日本経済新聞 電子版（日経電子版） @nikkei]

トヨタ、内製ヒト型ロボ「エリー」40万台を工場に　従業員と協働
https://t.co/3pG905strY

[Manda @Mandarobotics]

Astra is dominating existing robot policies. But what can Astra do that existing policies could not?

We analyzed 10 tasks, 5 Episodes each on Astra, Pi0.5, Nvidia’s Cosmos3 Nano and Gr00t 1.7 on RoboLab.

1) Astra is by far the most reliable model in terms of object recognition:

[LeRobot @LeRobotHF]

🌍 New World Action Model in LeRobot ✨

LaWAM (CoRL 2026) replaces future-video generation with a single-step latent subgoal.

Given an observation + language instruction, the policy predicts a latent action. A 230M Latent World Model decodes it into future visual features in ONE forward pass, and the action expert conditions on that predicted subgoal to produce the next chunk - dynamics-aware foresight, no iterative video rollout.

🎯 98.6% LIBERO · 91.22% RoboTwin · 90.0% real-world manipulation
⚡ 187ms / action chunk on A100, 10 denoising steps

Thanks for the contribution by Zhongguancun Academy 🙌
Docs: https://t.co/Dc59GaDllL
Blog: https://t.co/ETyQT38nfb

[Orboh_Kota Ueda(上田康太) @kotaueda999]

屋外の畑で、オクラを認知し、掴むことができました！カッターの製作が間に合わなくて挟むための割り箸をつけたけど、力���すぎてオクラ切断しちゃった笑
ランダムに生えてるオクラに対して、ロボットが対応し、ちゃんと掴みに行けるってすごいことだと思うし、屋外という難しい環境で作動したことがエンジニアとして嬉しかった。フィジ��ルAIの社会実装は今からでもできると実感したし、社会の役に立てると思いました。屋内の成功率は16/20です。開発日数は、70日です。約40日程度をUMIとDiffusion policy, Flow matching に費やしました。
開発で分かったことはこのnoteにまとめました。ぜひご覧ください。

https://t.co/PXiPIgVRfR

[Lingjun Mao @Lingjun_Mao]

wow!! GPT-6 Astra can rapidly adapt to a new robot embodiment and a new task 🤯🤖

We asked it to put a cake into an air fryer in the benchmark we’re building, using simple tool APIs including CuRobo-based motion planning.

Its first attempts were chaotic. After repeated interaction, it accumulated environment-specific experience and learned to reliably complete the task end to end. 🍰➡️🔥

This is generalization far beyond VLAs and WAMs.