Boston Dynamicsの秘密 The Secret of Boston Dynamics
しばらく見ないうちに、本家Boston Dynamicsの人型ロボットは進化し、バッテリー駆動の軽量になり、この中国製のロボットはそれにそっくりであった。Boston Dynamicsが最初に人型ロボットAtlasを発表した時、私はその滑らかで高速な動きに驚嘆した。それまでのロボットはASHIMOに代表されるように、動作がゆっくりでぎこちなかった。自由度は少なく、いかにも運動方程式を解いて動いている、基礎的なエンジニアリングの延長の域を出ていなかった。もちろんそのレベルでも、工場生産などの定型業務には十分役に立ったわけだが、汎用にするには越えなくてはいけない壁があった。While I hadn't been paying attention, Boston Dynamics' humanoid robots had evolved, becoming lightweight and battery-powered, and this Chinese-made robot closely resembled them. When Boston Dynamics first unveiled their humanoid robot, Atlas, I was astonished by its smooth and fast movements. Until then, robots, such as ASIMO, moved slowly and awkwardly. Their degrees of freedom were limited, and they clearly operated by solving motion equations, remaining within the realm of basic engineering. Of course, even at that level, they were sufficiently useful for routine tasks in factory production, but there was a significant hurdle to overcome for general-purpose use.
ここにNvidiaのIssacLabが作成したビデオが説明に使われており、これを見て、私は、以前から不思議に思ってきたBoston Dynamicsの人型ロボットの滑らかで高速な動作の仕組みがようやく理解できた。ニューラルネットを強化学習で利用しているのは想像していた通りだったが、これを高速の物理エンジン上で何億回も走らせることで学習させているようだ。翻って見るに、人間の学習は時間がかかる。赤ちゃんがハイハイを始めてからきちんと歩けるようになるまで、現実の物理エンジンを使って、実時間上でトレーニングを重ねていく。しかし、ロボットは仮想現実の物理シミュレーターの上で、何億回も瞬時に経験することが出来るため、学習速度が桁違いに早い。The video produced by Nvidia's IssacLab used for explanation helped me finally understand the mechanism behind the smooth and fast movements of Boston Dynamics' humanoid robots, which had long puzzled me. As I had imagined, they were using neural networks with reinforcement learning, but it seems they were training them by running millions of simulations on a high-speed physics engine. In contrast, human learning takes time. From the moment a baby starts crawling to when they can walk properly, they undergo training in real-time using the actual physical world. However, robots can experience millions of instances instantaneously in a virtual reality physics simulator, making their learning speed extraordinarily fast.
この桁違いの学習量はAIの進展の秘訣のようだ。DeepMindの２台のALPHA-ZEROは互いに何億回も対局の後世界一に。LLMは膨大なパラメータ数での学習。標題に上げたBoston Dynamicsのシミュレータ上での強化学習はAI進展の主流となっているし、今後もそのトレンドは続くだろう。This extraordinary amount of learning seems to be the secret behind AI advancement. DeepMind's two ALPHA-ZERO systems became the world's best after playing millions of matches against each other. Large Language Models (LLMs) learn with an enormous number of parameters. Reinforcement learning on simulators, as seen in Boston Dynamics' work highlighted in the title, has become the mainstream of AI advancement, and this trend will likely continue in the future.