Unverified50% confidenceFactExact time
Training a robot to walk in simulation allows it to experience tens of thousands of falls and get-ups at low cost, repeatably and at accelerated speed, whereas the corresponding process in the real physical world happens only once.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在方块世界中训练出的规划、导航、多步骤任务等抽象高层次能力比低层运动控制更容易跨域迁移到现实机器人和具身智能场景74% similarUnverifiedA generalist policy trained only on human demonstrations can complete locomotion tasks zero-shot on a real robot.73% similarUnverifiedOpenAI once trained a virtual robot to learn fast movement, and it discovered that stretching its body extremely thin and then falling over could move its center of mass farther, thereby achieving a high score.73% similarUnverified四足机器人的自我恢复能力来自强化学习的sim-to-real训练范式，通过在仿真环境中摔倒数百万次学习平衡策略72% similarUnverifiedIn reinforcement-learning locomotion research, if the reward function is designed too specifically, all robots converge to similar gaits; if it is too loose, training may fail to converge or learn odd behaviors that exploit loopholes.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1090302API
curl https://kongchang.com/api/v1/knowledge/claims/1090302MCP
get_claim(id=1090302)