Nikita Rudin discusses the gap between current robotics capabilities and real-world deployment, arguing that no humanoid robot today generates true value due to task mismatch.
要点 · TL;DR
机器人能去人类能去的任何地方才算解决运动问题。 Locomotion is unsolved until robots match human terrain traversal.
感知环节的仿真到现实差距更大，需精细模拟传感器。 Sim-to-real gap is larger with perception; careful sensor simulation needed.
当前人形机器人未创造真实价值，常需人类操作员。 Humanoid robots today lack real value; often need human handlers.
核心观点 · Key points
在机器人能可靠地到达人类能去的任何地方之前，运动能力问题尚未解决。 Locomotion is not solved until robots can go anywhere a human can, reliably.
加入感知后，仿真到现实的差距更大；需要仔细模拟传感器。 Sim-to-real gap is larger with perception; careful sensor simulation is needed.
模块化方法（分离运动与规划）在短期内是务实的。 Modular approach (separate locomotion and planner) is pragmatic short-term.
目前的人形机器人并未产生实际价值；它们通常需要人类操作员。 Humanoid robots today do not generate real value; they often need human handlers.
预训练的视觉语言模型有助于泛化，但动作头的训练仍是开放问题。 Pre-trained VLMs help generalization but action head training is still open.
价值将首先在工业场景中显现，然后才是消费场景，预计在2026年底。 Value will emerge in industrial settings first, then consumer, by late 2026.
反共识 · Contrarian takes
加入视觉因仿真到现实差距反而使运动更难，而非更易。 Adding vision makes locomotion harder due to sim-to-real gap, not easier.
端到端训练并非最佳方法；模块化系统更实用。 End-to-end training is not the best approach; modular systems are more practical.
人形机器人应以较低效率移动，以显得自然并建立信任。 Humanoids should move less efficiently to appear natural and build trust.
大多数令人印象深刻的机器人演示依赖于远程操作或大量数据收集。 Most impressive robot demos rely on teleoperation or massive data collection.
仿真并非总是更好；对于难以模拟的任务，需要真实数据。 Simulation is not always better; real data is needed for hard-to-simulate tasks.
人形并非必需；人类能力（双臂、移动性）才是关键。 Humanoid form is not essential; human capabilities (two arms, mobility) matter.
本期章节 · Chapters（共 26）
引言与Nikita背景Introduction and Nikita's background
演示与真实部署的差距The gap between demos and real-world deployment
人形机器人热点评论Hot take on humanoid robots
仿真到现实差距与运动挑战Sim-to-Real Gap and Locomotion Challenges
端到端与模块化方法End-to-End vs. Modular Approach
规划器目标与训练Planner Objective and Training
从四足到人形迁移Transfer from Quadrupeds to Humanoids
人形与四足仿真调优Simulation and Tuning for Humanoids vs Quadrupeds
人形机器人家庭就绪度Readiness of Humanoid Robots for the Home
机器人训练与数据收集挑战Challenges in Robot Training and Data Collection
理解仿真栈Understanding the Simulation Stack
切换机器人易，新任务难Switching Robots is Easy, New Tasks are Hard
分离模型与统一模型Separate vs. Unified Models
层级流水线与硬件约束Hierarchical Pipeline and Hardware Constraints
家庭机器人机载与云端计算Onboard vs. Cloud Compute for Home Robots
仿真环境与强化学习算法Simulation Environments and RL Algorithms
仿真中的抽象层次Levels of abstraction in simulation
结合模仿学习与强化学习Combining imitation learning and RL
仿真到现实差距与真实世界强化学习Sim-to-real gap and real-world RL
真实世界强化学习挑战与仿真优势Challenges in real-world RL and simulation benefits
奖励函数设计与价值函数Reward function design and value functions
未来几年机器人预测Predictions for robotics in the coming years
当前机器人硬件差异Hardware differences among current robots
硬件设计策略与竞争Hardware design strategies and competition
人形形态与人类能力Humanoid form factor vs. human capabilities