Gemini 4 Argon：机器人控制与代码生成提示示例
Gemini 4 Argon：机器人控制与代码生成提示示例
가이드 목록 · 본문은 주로 중국어 간체입니다. 영문 요약(English summary)을 참고하세요.
Gemini 4 Argon：机器人控制与代码生成提示示例
Gemini 4 Argon 是 Google DeepMind 在 2026 年 10 月 13 日发布的物理推理前沿模型，专门用于机器人控制、Real2Sim 重构和代码生成任务。它在 ERQA 基准上纯推理得分 87%，显著领先其他前沿模型。开发者可直接复制其开源提示模板，在 Cursor、Claude Code 或 Google AI Studio 中实现机器人任务自动化，特别适合需要视频驱动的移动操作、物理仿真和迭代优化的场景。工程师需根据项目复杂度选择模型，并结合自身订阅计划判断成本与集成路径。
机器人控制提示示例
Argon 擅长通过 MuJoCo 数字孪生练习、零样本控制和视频世界模型闭环调整实现多自由度操作。以下是官方代表性提示（直接复制粘贴使用）：
1. Spot 移动操作（垃圾桶踩踏扔垃圾）
Boston Dynamics Spot 四足机器人（12 自由度腿 + 6 自由度臂）学习踩脚踏垃圾桶丢垃圾。Argon 先创建脚踏垃圾桶的关节 MuJoCo 数字孪生练习步法，再切换到物理机器人。
You are an expert robotics control engineer. Control the Boston Dynamics Spot quadruped robot (12-DoF legs + 6-DoF arm) to learn a new skill: stepping on a foot-pedal trash bin to throw away soda cans. Create an articulated MuJoCo digital twin of the foot-pedal trash bin to practice the moves before stepping on the physical pedal and dropping the cans into the open bin. Use wrist camera feedback and closed-loop reasoning. Output the full action sequence with timestamps and MuJoCo simulation code for validation.
2. 零样本控制
简单机械臂将倒地马克杯竖起并将皱纸巾放入。Argon 使用视觉坐标约束和动作基元实现多轮操作。
Control a simple robotic arm to place a fallen mug upright and put the crumpled napkin inside. Use visual coordinate harness and action primitives over several turns. Provide the precise joint angle trajectories and vision-based feedback loop.
3. Omni 控制（视频世界模型闭环）
在 Gemini Omni World Model 中操控机器手将散乱的 Scrabble 拼字块（G、R、A、N、O）排列成 “A - R - G - O - N”。Argon 实时检查回滚关键帧并迭代优化提示。
Steer robots in the Gemini Omni World Model to rearrange five scattered Scrabble tiles (G, R, A, N, O) into a clean row spelling A - R - G - O - N. Output timestamped action prompts, inspect intermediate rollout keyframes, and refine the prompts in a closed visual loop.
Real2Sim 重构提示示例
Argon 可从真实机器人相机、人类视频或手绘草图重建物理校准的仿真孪生。
1. 机器人视图转仿真
从真实机器人遥操作片段生成 MuJoCo 数字孪生。
Given real robot teleoperation video, reconstruct a physically calibrated MuJoCo digital twin. Match sim visuals and contact events with real data in an iterative visual-reasoning loop. Output the full 3D scene assets, contact parameters, and synchronization code.
2. 人类视图转仿真
从人类操作视频生成 76 自由度双臂人形机器人孪生（替换包围盒、重新目标轨迹）。
From egocentric human manipulation video, identify task-relevant objects, replace bounding boxes with compound convex-decomposed meshes, and retarget the human hand trajectory onto a 76-DoF bimanual humanoid with five-fingered hands in MuJoCo. Provide the retargeted trajectory and multi-view render code.
3. 草图转仿真
手绘草图（阿波罗 2 人形投球弧线击倒 3-2-1 盒子堆）自动生成 MuJoCo_warp 场景和平衡抛射轨迹。
From a hand-drawn sketch of an Apollo 2 humanoid throwing a ball along a parabolic arc to topple a 3-2-1 stack of boxes, automatically construct the GPU-accelerated MuJoCo_warp scene, synthesize balanced 76-DoF whole-body trajectory, verify 6/6 knockdown, and export Blender 4.5 Cycles render.
物理理解与自改进提示示例
Argon 提供多粒度 hindsight 标注、值函数和模拟迭代优化。
1. hindsight 标注与数据质量
对 ALOHA 2 机器人折纸狐狸视频生成 4 级标注层级（目标、子任务、原子动作、QC）。
Analyze the overhead video of a bimanual ALOHA 2 robot folding a square sheet of paper into an origami fox over 136 seconds. Produce a time-aligned 4-level hindsight annotation hierarchy: Episode Goal (G1), 6 contiguous Semantic Subtasks (G2), 17 contact-level Atomic Primitives (G3), and automated Quality Control (QC) reweighting.
2. 值函数
评估 60 秒双臂环形分拣轨迹，生成密集值函数和 28 段进度斜率。
Generate dense value function / task progress predictions for the synchronized head, left-gripper, and right-gripper camera feeds around contact transitions. Decompose the trajectory into 28 segments with positive, neutral, and negative progress slopes.
3. 模拟驱动自改进
从单张 RGB 照片构建模拟孪生，合成 4500+ 域随机化片段提升真实成功率（56% → 75%）。
Starting from a single RGB photo of the real tabletop, build a calibrated MuJoCo simulation twin, define new tasks, and synthesize a closed-loop motion generator to collect 4,500+ domain-randomized episodes. Boost real-world success rate and reduce completion time.
提示使用建议
将以上模板输入 Cursor（支持多模态上下文）或 Google AI Studio 的 gemini-4-argon 模型中。提示长度建议 200–500 token 以获得稳定输出。结合 MuJoCo 引擎和 Google SDK 可快速验证。实际成本参考 Google AI Studio 官方定价页（以当日挂牌为准）。
Gemini 4 Argon 与主流 AI 编码工具对比
| 特性 | Gemini 4 Argon | Cursor / Claude Code | OpenAI / Grok |
|---|---|---|---|
| 机器人控制能力 | 高（MuJoCo 闭环 + 视频世界模型） | 中（代码生成为主） | 中（工具调用有限） |
| Real2Sim 重构 | 原生视频 + 草图转仿真 | 需辅助脚本 | 需外部工具 |
| 物理理解标注 | 4 级 hindsight + 值函数 | 文本描述 | 文本描述 |
| API 集成易用性 | 原生 Google SDK | Cursor API 包 | OpenAI / Grok API |
| 典型 Token 成本 | $2 / $10 / 1M（输入/输出） | 订阅制（不公开单价） | 订阅制（不公开单价） |
风险与边界
此指南仅为技术参考，不构成任何法律意见或安全建议。使用提示时请评估机器人硬件安全、数据隐私和算法偏差。实际效果取决于具体机器人平台与场景，建议在仿真环境中验证后再上物理平台。Google 保留对模型和定价的最终解释权，以官方页面为准。
延伸阅读
English summary
Gemini 4 Argon is Google DeepMind’s new frontier physical reasoning model released on October 13, 2026. It excels at robot control, Real2Sim reconstruction, and physical understanding using MuJoCo simulation and video inputs. This guide provides ready-to-copy prompts for mobile manipulation, zero-shot control, video world model steering, sketch-to-simulation, hindsight labeling, value functions, and self-improvement loops. These templates integrate directly with Cursor, Claude Code, or Google AI Studio for embodied AI workflows. Pricing is $2/$10 per 1M tokens (input/output) for the Gemini API tier. Compare with other tools in the table above. Always validate in simulation first and check official pricing pages for current rates. This content is for developer reference only.