Model action and consequence together.
统一建模行动与后果。
The same representation should support acting in the world and imagining how the world will change.
让同一套表征既能决定怎么做，也能预测做了以后世界会如何变化。
Physical intelligence, grounded in consequence以因果后果为基础的物理智能
Intelligence
through consequence.
让机器人，
从后果中学习。
We are building foundation models for robots that understand the physical world, choose actions, anticipate what happens next, and improve from every interaction.
我们正在构建机器人基础模型：理解物理世界、选择行动、预见结果，并从每一次真实交互中持续进化。
WORLD × ACTIONSHARED MODEL
01 OBSERVE
02 ACT
03 PREDICT
04 IMPROVE
01 / THESIS
Robots need more than action prediction.
机器人需要的，不只是动作预测。
A capable robot must connect perception, action, consequence, and value inside one learning system. Our work is organized around three ideas.
真正通用的机器人，需要在同一个学习系统中连接感知、行动、后果与价值。我们的技术路线围绕三个核心判断展开。
Learn from heterogeneous experience.
从异构经验中学习。
Human video, robot demonstrations, simulation, and autonomous rollouts each reveal a different part of physical intelligence.
人类视频、机器人示范、仿真与自主执行，共同提供关于物理智能的不同监督。
Turn deployment into learning.
让部署本身成为学习过程。
Failures, corrections, and recoveries are not edge cases to discard. They are the highest-value data for the next model.
失败、纠正与恢复不应被丢弃，它们是下一代模型最有价值的数据。
02 / ARCHITECTURE
One model. Two modes. One learning loop.
一个模型，两种模式，一个闭环。
A shared world–action backbone connects policy learning with simulation. What the model learns about the future can improve how it acts; what it experiences while acting can improve its model of the world.
共享的 World–Action Backbone 连接策略学习与仿真。对未来的理解改善行动，真实行动产生的经验又反过来更新世界模型。
UNIFIED WORLD–ACTION MODEL / CONCEPTUAL ARCHITECTURE
Visual history视觉历史o(t−k:t)
●Robot state机器人状态s(t−k:t)
◆Language intent语言意图instruction
⌁
SHARED BACKBONE
Physical representation
for prediction and control 面向预测与控制的
统一物理表征
for prediction and control 面向预测与控制的
统一物理表征
A common latent space aligns what the robot sees, what it can do, and what happens next.
在同一个潜在空间中，对齐机器人看到的世界、可执行的动作与未来后果。
ACTION MODE行动模式
Predict executable robot actions and future observations.
预测可执行动作与未来观测。
SIMULATION MODE仿真模式
Predict consequences conditioned on proposed actions.
基于候选动作预测世界后果。
LEARNING LOOP学习闭环
Generate recovery experience and improve closed-loop behavior.
生成恢复经验，持续提升闭环表现。
03 / DATA FLYWHEEL
Every failure can become capability.
每一次失败，都可以变成能力。
Real-world rollout reveals what offline data cannot. We turn deviations, interventions, and recoveries into structured supervision—then feed them back into the shared model.
真实执行会暴露离线数据看不到的问题。我们把偏差、人工介入与恢复过程转化为结构化监督，再送回统一模型。
04 / SYSTEM
Intelligence at the right timescale.
让不同层级，在正确的时间尺度上思考。
General-purpose autonomy needs deliberation, embodied skill, and responsive control to work as one system—without forcing one model to do everything.
通用自主性需要任务推理、具身技能与实时控制协同工作，而不是让一个模型勉强承担所有职责。
AGENTSECONDS → MINUTES
Reason, plan, recover.
推理、规划与恢复。
Understands intent, decomposes tasks, verifies outcomes, and replans when the world does not match expectations.
理解意图、拆解任务、验证结果，并在现实偏离预期时重新规划。
VLAHUNDREDS OF MS
See, generalize, manipulate.
感知、泛化与操作。
Connects visual-language understanding with diverse robot skills and long-horizon manipulation.
连接视觉语言理解、多样化机器人技能与长时序操作。
SONICMILLISECONDS
React, stabilize, execute.
响应、稳定与执行。
Provides fast, robust control for contact-rich motion and precise physical execution.
面向高接触、高动态任务，提供快速、稳定而精确的物理执行。
Our direction我们的方向
Build robots that grow more capable with every experience.
构建能够随经验持续成长的机器人智能。
Causa Robotics is assembling a founding team in Shanghai across robot learning, world models, control, data infrastructure, and system integration.
灵知机器人正在上海组建创始团队，方向覆盖机器人学习、世界模型、运控、数据基础设施与系统集成。
SHANGHAI · 2026 →