🇯🇵🇮🇹夫婦 / 機械 → EV&ロボティクス → M&A @ 駐在
Japan, Switzerland, Italy, USA
Joined March 2008
- Tweets19.4K
- Following293
- Followers401
- Likes5.7K
chachay retweeted
フィジカルAIの領域でかなり革新的な成果が出てきました。現在、ヒューマノイドの全身制御（motion tracking）ではsim-to-realによる強化学習が活発に行われていますが、ここで使われている物理演算シミュレータを、ニューラルネットベースの世界モデル（world model）に置き換えることで、sim-to-realを1000倍近く高速化かつ洗練できるという技術です。
xtech.nikkei.com/atcl/nxt/ma…
開発したのはロボット機械学習の領域の世界的権威ともいえるPieter Abbeel氏（米University of California Berkeley 教授）の研究グループ。今週、米国Austinで開催中の #CoRL2026 でちょうど昨晩oralによる発表がありました。ダイナミクスの模擬を微分不可能な従来型のシミュレータではなく、ニューラルネットベースにしてやることで目的関数→reward→world model→policyと全域が微分可能になり、通常の勾配法で学習できることも、policyの洗練度合いに寄与しているようです。日経Robotics 最新号 12月号でこの技術「SuperMimic」について10ページにわたってその仕組み・インパクトを解説した記事を書きました。
この領域ではNVIDIAのSONICが記念碑的成果になっていましたが、もはやSONICが古くさく見えてしまうほど洗練された成果だと感じます。SONICでは失敗していた後方宙返りなどのアクロバティックな動作もSuperMimicでは成功させています。
それと、10月の論文発表段階ではあまり詳細が明らかになっていませんでしたが、CoRL2026の講演では、冗談っぽく「Top Secret」として論文投稿後の進展についても言及があり、やはり単なる全身制御だけでなく、既にマニピュレーションにも適用していて良い成果が出つつあることも明かされていました。多指ハンドでもPPO利用のケースを大幅に上回るようです。CoRL2026での講演と同時にコードもリリースされました。SuperMimicは世界的にはかなり注目されていて、コードがリリースされる以前から、後続研究が次々と出ています。まさに、sim-to-realに革新を起こしつつあるうねりを感じます。
chachay retweeted
5,000 years of world history on a single timeline completely interactive and explorable.
entireworldhistorytimeline.c…
chachay retweeted
Even better an interactive website entireworldhistorytimeline.c…
chachay retweeted
5,000 years of world history on a single timeline. Zoom in. You could spend 30 minutes exploring this.
chachay retweeted
海外勤務者の過労死相次ぐ 厚生労働省、労働実態調査へ | 毎日新聞 mainichi.jp/articles/2026110…
chachay retweeted
Can reinforcement learning replace traditional inverse kinematics for robot arms?
PolicyIK takes an interesting approach.
Instead of relying on numerical IK solvers and hand-tuned controllers, it trains a goal-conditioned RL policy to convert end-effector pose errors into joint-position updates.
The team also trained the policy directly on a physical robot and reported better reaching and holding performance than four classical IK controllers.
This could make the precision-control layer behind VLA-based robot manipulation more adaptable.
Can reinforcement learning solve inverse kinematics precisely?
Introducing 𝗣𝗼𝗹𝗶𝗰𝘆𝗜𝗞 🦾, a goal-conditioned RL policy for high-precision inverse kinematics.
On a real arm, it reaches median errors of 𝟬.𝟱𝟬 𝗺𝗺 𝗮𝗻𝗱 𝟬.𝟭𝟮° and holds the goal pose on 𝟮.𝟱× 𝗺𝗼𝗿𝗲 𝘁𝗮𝘀𝗸𝘀 than the best classic IK controller.
🌐 tinaliu0123.github.io/Policy…
🧵👇
chachay retweeted
Each of the 400k particles is a unique, untrained, randomly mutated ~10k param neural net btw 🤗
🔊🦠 Neural-Kimchi-4B 🔊🦠
chachay retweeted
NEW SIMULATION: Keeping track of your sims actually helps better than running more of them!
Called SGS, or Success Guided Sampling, it is a project from the University of Washington and NVIDIA.
It is standard PPO with one extra loop around it.
Every task configuration, a tuple of initial state, goal and environment, keeps a circular buffer of its last H=100 Boolean outcomes; the mean of that buffer is its success rate p_i.
Configurations are then sampled with weight w_i, a beta-shaped kernel in mode-concentration form, pushed through a softmax at temperature T, so that whatever the policy currently succeeds at roughly t of the time gets sampled most.
Locomotion uses over 104,000 configurations built from 13 terrain types, 10 difficulty levels, 20 start states and 20 goals.
Training runs in MuJoCo on L40S and H200 GPUs at up to 32,768 and 16,384 environments per GPU.
Robots are ANYmal C and D, a Franka Panda and a UR5e.
-> The contribution is roughly bookkeeping.
SGS keeps a 100-entry Boolean ring buffer per configuration, computes a mean, evaluates a two-term product, and softmaxes.
No new objective, no architecture, no auxiliary network, no demonstrations, no reward shaping.
chachay retweeted
We trained robot policies in different physics engines (MuJoCo, PhysX, Newton, & Genesis) and tested how well they transfer.
Even with identical assets, the choice of engine changes learned behavior and success rates.
1. The same policy might fail out of its original engine 🧵
chachay retweeted
The Jacobian matrix holds every first partial derivative of a map, so it shows how a small change in each parameter moves the output point.
Here (s, t, h) maps to the point at distance h along the ray through (s, t, 1), with w = √(1 + s² + t²). Each row is a tangent vector: t_s and t_t run along the grid lines of the sphere, t_h points straight out from the centre.
The cross product t_s × t_t gives the surface normal, and the determinant h²/w³ is the volume factor that every change of coordinates in a multiple integral needs.
chachay retweeted
"cuArms: GPU-Accelerated Conflict-Based Planning for Multi-Arm Manipulation"
cuarms.github.io/
From Nvidia. Really happy to see that PIBT+LaCAM can now solve the multi-arm coordination problem so effectively!
chachay retweeted
I joined @GoogleDeepMind to develop a research program on the economics of AGI. Today, together with Andrew Koh (@andrewjkoh) and Shane Legg (@ShaneLegg), we are incredibly excited to release The Economics of Intelligence, the first in a series of projects from our group on how AGI will change the economy.
How do capability improvements transform the economy, and when? In this project, we define thresholds of economic capabilities, e.g., minimal AGI, based on how they transform the economy. We then track how the economy–market structure, employment, and wages–changes during the transition across capability regimes.
We have seen exponential increases in what AI agents can do but the broader economy seems almost remarkably unchanged. Part of the reason is that the solo AI agent is the wrong economic object; economic transformation is an equilibrium outcome where organizations compete. Looking at how intelligence functions within organizations, transformation will depend on when two key capability thresholds are crossed: competence (gains on individual skills, e.g., coding, math) and coordination (gains on ability of agents to cooperate and work together).
We show that what economic transformation looks like depends on whether AI becomes supercompetent, supercooperative, or both—the regime of artificial superintelligence (ASI). Tracking this transition shows why the real visible transformation may come from improvements in coordination—a capability most benchmarks don’t yet track. Past a certain threshold, AI capabilities can break the Coasian ceiling on firm size and lead to an economy that looks drastically different than the one we see today.
Two main takeaways:
1. To understand the path of transformation, we should be tracking agentic coordination more closely.
2. The impact of AI will likely show up in wages before market structure or employment.
We are really excited about this paper. Excited to hear people’s thoughts.
Link: institute.deepmind.com/essay…
chachay retweeted
仮定を鍵、結論を扉と見立てて、ダンジョン探索が「証明」になっていくブラウザゲーム「Proof Dungeon: 自然演繹の迷宮」を作りました。パズル感覚で論理学や自然演繹を学べるブラウザゲームです。日本語と英語に対応しています。
kihara-logic.github.io/proof…
制御工学っぽい動画だけど、他の分野との接続があるんだろうか
A 2×2 matrix A makes the plane flow by x′ = Ax, and each flow is one of nine kinds: points sink, escape, spiral, circle or settle on a line.
The kind is set by the eigenvalues of A, the roots of λ² − (tr A)λ + det A = 0. det A < 0 gives a saddle. Above det A = (tr A)²/4 they are complex: spirals, or a centre if tr A = 0. Below it, nodes; on it, star and degenerate node.
The stable spiral here has tr A = −1, det A ≈ 2.65, so λ = −0.5 ± 1.55i.
62
chachay retweeted
トークナイザーを開発している奇特な人にしか需要ないだろうけど、各種LLMのトークン化をブラウザで実験できるページを作成しておいた imoz.jp/tokenizer/index.html
