Google DeepMind演示Gemini 4控制人形机器人玩投掷游戏
Teaching a humanoid robot to play cornhole with Gemini 4 Argon
展示了Gemini 4在多模态理解与具身智能控制上的实际落地能力，对关注Agent和机器人领域的开发者有参考价值。
All right.
好的。
In this experiment, we are basically trying to see if we can teach Gemini 4 to control a humanoid robot to solve cornhole. We just showed it a video of a human doing the task and we prompted Gemini 4 to run the Mojoku simulator and like practice and then we let it learn from experience, iterate, make mistakes and look at each video of what went wrong. And with experience of each trial, it's able to improve what it can do better.
在这个实验中，我们基本上是想看看能否教 Gemini 4 控制人形机器人来解决 cornhole（沙包入袋）游戏。我们只是给它看了一段人类执行该任务的视频，并提示 Gemini 4 运行 Mojoku 模拟器进行练习，然后让它从经验中学习、迭代、犯错，并查看每次出错的视频。凭借每次尝试的经验，它能够改进自己的表现。
Woohoo! Airmail straight in the hole. Let's go. To sink those cornhole shots, I used four steps. First, I watched a video of a person throwing a bean bag, tracking their arm swing and release to map that human motion onto my own robot joints and hands. Second, I looked through my two eye cameras to measure the exact 3D position of the board and built a physics simulation of my body and the board. Third, I practiced the throwin simulation, widening my stance and tuning my swing so my hand cleanly cleared my hip and launched a smooth arc.
太棒了！空中传球直接进洞。走起！为了完成那些 cornhole 投掷，我用了四个步骤。第一，我观看了一段有人扔沙包的视频，追踪其手臂摆动和释放动作，将这些人类运动映射到我的机器人关节和手上。第二，我通过我的双目摄像头测量了目标板的精确三维位置，并构建了我和目标板的物理仿真模型。第三，我在仿真中练习投掷，加宽站距并调整摆动幅度，让我的手干净利落地避开髋部，并划出平滑的弧线。
And fourth, on the real robot, I watched each bag land and adjusted my release by a few cm until we sank seven shots straight in the hole. Woohoo! Gemini 4 is able to reason through its video understanding and narrow down to like what went wrong and it was learning from its own experience which was extremely thrilling to see.
第四，在真实机器人上，我观察每个沙包的落点，并通过几厘米的调整来修正释放时机，直到我们连续七次成功投中。太棒了！Gemini 4 能够基于其视频理解能力进行推理，缩小问题范围，找出哪里出了错，并且它是在从自身经验中学习，这看起来令人极度兴奋。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力