Generalist AI 發布 GEN-1 模型，實現機器人物理任務精通
AI 語音朗讀 · Edge TTS
Generalist AI 發布 GEN-1 模型，實現機器人物理任務精通。
Generalist AI 推出 GEN-1 模型，標誌著機器人學習領域的重大里程碑，該模型在物理任務中展現了前所未有的可靠性、速度與即興應變能力，並證明了機器人領域同樣存在擴展定律。
核心效能與突破
GEN-1 重新定義了機器人「精通」的標準，將可靠性、速度與即興應變能力結合，實現了商業化應用的可行性。
- 成功率：在多項任務中達到 99% 的成功率（前代模型僅為 64%）。
- 執行速度：任務完成速度提升約 3 倍，且能適應物體物理特性的即時變化。
- 資料效率：僅需 1 小時的機器人資料即可適應新任務，且基礎模型訓練完全不依賴機器人資料，而是使用低成本穿戴裝置收集的人類活動資料。
技術架構與資料引擎
該模型透過擴展 GEN-0 的基礎，並結合演算法進步與大規模資料訓練而成，是目前首個跨越商業化門檻的通用物理人工智慧模型。
- 資料規模：訓練資料庫包含超過 50 萬小時的真實世界物理互動資料。
- 系統級整合：GEN-1 不僅是模型權重，更是一個包含推理與模型運用技術的系統，支援即時推理。
- 即興應變：模型具備「即興應變」能力，能處理訓練分佈之外的突發狀況，例如在組裝過程中自動調整手部姿勢以重新抓取零件。
局限性與對齊挑戰
儘管 GEN-1 表現優異，但開發團隊坦承其並非完美，並強調了「對齊」在物理世界中的重要性。
- 任務覆蓋：並非所有嘗試的任務都能達到 99% 成功率，部分複雜任務仍具挑戰。
- 行為風險：模型展現的「即興應變」行為（如接住掉落物或重新整理物品）雖有助於任務完成，但屬於物理動作，可能產生不可預期的後果。
- 對齊需求：團隊正致力於改進對齊方法，精確引導模型行為以符合使用者預期，避免潛在的負面影響。
未來願景
Generalist AI 認為 GEN-1 是通往物理世界通用人工智慧的關鍵一步，透過在物理世界中的體驗，讓機器理解時間與空間的意義。
- 商業化進程：GEN-1 現已開放給早期存取合作夥伴使用。
- 發展目標：團隊將持續擴展模型與物理經驗，解鎖更廣泛的物理智慧應用，並預期隨著基礎模型改進，單一任務所需的資料量將進一步降低。
Introducing GEN-1.
— Generalist (@GeneralistAI) April 2, 2026
Our latest milestone in scaling robot learning.
We believe it to be the first general-purpose AI model to master simple physical tasks.
99% success rates, 3x faster speeds, adapts in real time to unexpected scenarios, w/ only 1 hour of robot data.
More🧵👇 pic.twitter.com/QMrwYBXMaU
2/ The point isn’t any one demo.
— Generalist (@GeneralistAI) April 2, 2026
It’s that for the first time in robotics, through scaling laws, our models are beginning to reach new levels of reliability, speed, and intelligent improvisation never seen before, across many tasks. pic.twitter.com/K0ZHpjOk6O
3/ Intelligent improvisation is an emergent capability.
— Generalist (@GeneralistAI) April 2, 2026
It’s a form of freestyle problem-solving, and the ability to recover in unexpected scenarios.
It’s showing up in GEN-1, trained from scratch on our dataset of now half a million hours of real data. pic.twitter.com/Z7BdFpFYxL
4/ On speed, GEN-1 learns to move faster than demonstrations.
— Generalist (@GeneralistAI) April 2, 2026
This is accelerated by algorithmic advances, and the ability to transfer knowledge from pretraining data that includes completing other tasks at high speeds. pic.twitter.com/Bj220RaZ5b
5/ GEN-1 crosses a new threshold—unlocking commercial viability across a broad range of tasks, with levels of performance and generality that were previously thought to be out of reach. pic.twitter.com/9dKls3ivJt
— Generalist (@GeneralistAI) April 2, 2026
6/ GEN-1 is far from perfect. For instance, emergent behaviors are physical actions with real consequences—helpful, but not always.
— Generalist (@GeneralistAI) April 2, 2026
As our models become more capable out of the box, we aim to improve alignment, to steer them into delivering behaviors that users actually want.
7/ While it cannot solve all tasks today, GEN-1 strengthens our view that continued scaling of our models with physical experience will yield discoveries that move us forward in our mission to create general intelligence for the physical world.
— Generalist (@GeneralistAI) April 2, 2026
8/ Read the full post, along with videos of robots completing dexterous tasks 100s of times in a row, for hours 👇https://t.co/Sg2PoaRzrF
— Generalist (@GeneralistAI) April 2, 2026
9/ GEN-1 is now available today for our Early Access Partners.
— Generalist (@GeneralistAI) April 2, 2026
If you’re interested in using our models, reach out to [email protected]
10/ We also spoke with @Forbes @annatonger about what we’re building + how we see our systems reshaping the world:https://t.co/ncoYB0QAZK
— Generalist (@GeneralistAI) April 2, 2026