2026年要关注的人工智能机器人公司
2026年最重要的人工智能机器人公司:图像,物理智能,Covariant,Unitree,以及新兴的创业公司,
2026 年的景观
智能化机器人行业进入了一个新阶段.经过多年的渐进进,基础模型,大规模示范数据和负担得起的硬件的融合正在产生可以在任务中通用的机器人. 2025年,人工智能机器人投资超过80亿美元,仅人类型公司筹集了超过30亿美元. 本指南概述了四个类别最重要的公司:人工智能/基础模型初创公司,人形机器人公司,工业机器人创新者和技术提供商.
在RCSV,我们跟踪这些公司,因为他们的产品和研究直接塑造了我们的客户需要的硬件和数据.下列列许多机器人和平台可在我们的 [旧金山和奥尔斯顿实验室]测试.
实体化人工智能和基础模型公司
这些公司正在构建"大脑"-- 一般用途的AI模型,可以控制各种机器人体在许多任务中.
| Company |
HQ |
Funding |
Approach |
Why It Matters |
| Physical Intelligence (Pi) |
San Francisco |
$400M+ |
VLAs (vision-language-action models) trained on massive multi-robot datasets |
If Pi succeeds, a single model controls any robot body -- collapsing the cost of robot programming |
| Covariant |
Emeryville, CA |
$222M |
RFM-1 foundation model for robotic manipulation; production-deployed picking |
Most production-validated AI manipulation company; proving foundation models work in logistics |
| Skild AI |
Pittsburgh |
$300M+ |
General-purpose robot brain; cross-embodiment model |
CMU pedigree; betting on a single model for arms, quadrupeds, and humanoids |
| Octo / RT-X (Google DeepMind) |
San Francisco |
Internal (Alphabet) |
Open-source cross-embodiment models trained on Open X-Embodiment dataset |
Sets the open-source baseline; their dataset standard (RLDS) is becoming the lingua franca |
| Dobb-E / NYU RAIL |
New York |
Research grants |
Home robot learning from iPhone demonstrations; low-cost hardware |
Demonstrates that useful manipulation can come from consumer-grade data collection |
人类机器人公司
| Company |
HQ |
Funding |
Robot |
Why It Matters |
| Figure |
Sunnyvale, CA |
$750M+ |
Figure 02 -- 5'6", ~60 kg, dexterous hands, BMW factory deployment |
First humanoid with real factory deployment; partnership with OpenAI for language-grounded control |
| Unitree Robotics |
Hangzhou, China |
$150M+ |
G1 humanoid ($16K starting), H1 ($90K), Go2 quadruped |
Price disruption -- G1 at $16K makes humanoid research accessible to universities. RCSV stocks the G1 for lease |
| Tesla Optimus |
San Francisco, CA |
Internal (Tesla) |
Optimus Gen 2 -- 5'8", 57 kg, 11-DOF hands |
Manufacturing scale potential via Tesla's factories; FSD neural net team working on manipulation |
| 1X Technologies |
Moss, Norway |
$225M+ |
NEO Beta -- bipedal, soft actuators, designed for homes |
Unique soft-actuator approach for safety; backed by OpenAI Startup Fund |
| Apptronik |
Austin, TX |
$350M+ |
Apollo -- 5'8", 73 kg, 25 kg payload, partnership with Mercedes-Benz |
NASA heritage (Valkyrie lineage); focused on logistics and automotive manufacturing |
| Agility Robotics |
Corvallis, OR |
$200M+ |
Digit -- bipedal, 16 kg payload, Amazon warehouse pilot |
First humanoid with a dedicated factory (RoboFab); closest to volume production |
工业机器人创新者
| Company |
HQ |
Focus |
Why It Matters |
| Machina Labs |
Chatsworth, CA |
AI-driven sheet metal forming with industrial robots |
Applying AI to transform traditional manufacturing; Boeing and Lockheed contracts |
| Formic |
Chicago, IL |
Robots-as-a-Service (RaaS) for SMB manufacturers |
$0 upfront, pay-per-hour model makes automation accessible to small factories |
| Realtime Robotics |
Boston, MA |
Hardware-accelerated collision-free motion planning |
Replaces MoveIt-style planning with FPGA-based <1ms path generation; enables multi-robot cells |
| Flexiv |
Santa Clara / Shanghai |
Adaptive force-controlled robots (Rizon series) |
Best-in-class force control for contact-rich assembly tasks |
技术公司的发展
| Company |
Focus |
Why It Matters |
| NVIDIA (Isaac) |
GPU-accelerated simulation (Isaac Sim), inference (Jetson), robot foundation models (GR00T) |
The default simulation and training infrastructure for AI robotics companies |
| Genesis (simulation) |
Open-source physics simulator; GPU-accelerated; MuJoCo-compatible |
Fastest growing open-source sim; enables massive parallel RL training |
| Foxglove |
Robot data visualization and fleet management platform |
Replacing RViz for production robots; the "Datadog for robotics" |
| Paxini (RCSV partner) |
High-resolution tactile sensors for dexterous manipulation |
Tactile sensing is the missing modality; available at RCSV |
| Sanctuary AI |
Carbon robot hand with 20-DOF; AI work system |
Most advanced dexterous hand on a general-purpose humanoid |
资金的风景
机器人风险投资集中在三个领域:
- **人类:**数字 ($750M+),Apptronik ($350M+),1X ($225M+),敏捷 ($200M+).这些公司是前收入或早期收入,估值是由一般劳动力巨大的TAM驱动的.
- **操纵的基础模型:**物理智能 ($400M+),技能 ($300M+),共变 ($222M).投注是机器人AI遵循LLM扩展的玩法 - - 更多数据和计算产生更有能力的模型.
- 硬件启用者: 单位 (约合150亿美元以上),Flexiv (约合100亿美元以上).这些公司今天向真正的客户出售真正的硬件,并随着人工智能繁荣增长的收入.
对于进入这个领域的初创公司,我们建议专注于垂直应用 (特定行业的特定任务) 而不是竞争于一般用途的AI或硬件.
对于田野的成功意味着什么
- **如果Pi/Covariant基础模型成功:**机器人编程成为快速工程.瓶完全转向硬件质量和数据收集 - - 正是RCSV数据服务提供的.
- **如果Figure/Apptronik在规模上部署:**人体进入物流和制造业,从而产生了对培训数据,远程操作系统和维护基础设施的巨大需求.
- **如果Unitree的G1扩大:**大学实验室获得负担得起的人类型平台,加速研究.RCSV已经支持G1在我们的 租计划 和 数据平台.
- **如果开源 (Octo,Genesis,LeRobot) 获胜:**进入障碍下降,更多的初创公司进入,以及对质量数据和可靠硬件的需求增加.
2026 年的景观
智能化机器人行业进入了一个新阶段.经过多年的渐进进,基础模型,大规模示范数据和负担得起的硬件的融合正在产生可以在任务中通用的机器人. 2025年,人工智能机器人投资超过80亿美元,仅人类型公司筹集了超过30亿美元. 本指南概述了四个类别最重要的公司:人工智能/基础模型初创公司,人形机器人公司,工业机器人创新者和技术提供商.
在RCSV,我们跟踪这些公司,因为他们的产品和研究直接塑造了我们的客户需要的硬件和数据.下列列许多机器人和平台可在我们的 [旧金山和奥尔斯顿实验室]测试.
实体化人工智能和基础模型公司
这些公司正在构建"大脑"-- 一般用途的AI模型,可以控制各种机器人体在许多任务中.
| Company |
HQ |
Funding |
Approach |
Why It Matters |
| Physical Intelligence (Pi) |
San Francisco |
$400M+ |
VLAs (vision-language-action models) trained on massive multi-robot datasets |
If Pi succeeds, a single model controls any robot body -- collapsing the cost of robot programming |
| Covariant |
Emeryville, CA |
$222M |
RFM-1 foundation model for robotic manipulation; production-deployed picking |
Most production-validated AI manipulation company; proving foundation models work in logistics |
| Skild AI |
Pittsburgh |
$300M+ |
General-purpose robot brain; cross-embodiment model |
CMU pedigree; betting on a single model for arms, quadrupeds, and humanoids |
| Octo / RT-X (Google DeepMind) |
San Francisco |
Internal (Alphabet) |
Open-source cross-embodiment models trained on Open X-Embodiment dataset |
Sets the open-source baseline; their dataset standard (RLDS) is becoming the lingua franca |
| Dobb-E / NYU RAIL |
New York |
Research grants |
Home robot learning from iPhone demonstrations; low-cost hardware |
Demonstrates that useful manipulation can come from consumer-grade data collection |
人类机器人公司
| Company |
HQ |
Funding |
Robot |
Why It Matters |
| Figure |
Sunnyvale, CA |
$750M+ |
Figure 02 -- 5'6", ~60 kg, dexterous hands, BMW factory deployment |
First humanoid with real factory deployment; partnership with OpenAI for language-grounded control |
| Unitree Robotics |
Hangzhou, China |
$150M+ |
G1 humanoid ($16K starting), H1 ($90K), Go2 quadruped |
Price disruption -- G1 at $16K makes humanoid research accessible to universities. RCSV stocks the G1 for lease |
| Tesla Optimus |
San Francisco, CA |
Internal (Tesla) |
Optimus Gen 2 -- 5'8", 57 kg, 11-DOF hands |
Manufacturing scale potential via Tesla's factories; FSD neural net team working on manipulation |
| 1X Technologies |
Moss, Norway |
$225M+ |
NEO Beta -- bipedal, soft actuators, designed for homes |
Unique soft-actuator approach for safety; backed by OpenAI Startup Fund |
| Apptronik |
Austin, TX |
$350M+ |
Apollo -- 5'8", 73 kg, 25 kg payload, partnership with Mercedes-Benz |
NASA heritage (Valkyrie lineage); focused on logistics and automotive manufacturing |
| Agility Robotics |
Corvallis, OR |
$200M+ |
Digit -- bipedal, 16 kg payload, Amazon warehouse pilot |
First humanoid with a dedicated factory (RoboFab); closest to volume production |
工业机器人创新者
| Company |
HQ |
Focus |
Why It Matters |
| Machina Labs |
Chatsworth, CA |
AI-driven sheet metal forming with industrial robots |
Applying AI to transform traditional manufacturing; Boeing and Lockheed contracts |
| Formic |
Chicago, IL |
Robots-as-a-Service (RaaS) for SMB manufacturers |
$0 upfront, pay-per-hour model makes automation accessible to small factories |
| Realtime Robotics |
Boston, MA |
Hardware-accelerated collision-free motion planning |
Replaces MoveIt-style planning with FPGA-based <1ms path generation; enables multi-robot cells |
| Flexiv |
Santa Clara / Shanghai |
Adaptive force-controlled robots (Rizon series) |
Best-in-class force control for contact-rich assembly tasks |
技术公司的发展
| Company |
Focus |
Why It Matters |
| NVIDIA (Isaac) |
GPU-accelerated simulation (Isaac Sim), inference (Jetson), robot foundation models (GR00T) |
The default simulation and training infrastructure for AI robotics companies |
| Genesis (simulation) |
Open-source physics simulator; GPU-accelerated; MuJoCo-compatible |
Fastest growing open-source sim; enables massive parallel RL training |
| Foxglove |
Robot data visualization and fleet management platform |
Replacing RViz for production robots; the "Datadog for robotics" |
| Paxini (RCSV partner) |
High-resolution tactile sensors for dexterous manipulation |
Tactile sensing is the missing modality; available at RCSV |
| Sanctuary AI |
Carbon robot hand with 20-DOF; AI work system |
Most advanced dexterous hand on a general-purpose humanoid |
资金的风景
机器人风险投资集中在三个领域:
- **人类:**数字 ($750M+),Apptronik ($350M+),1X ($225M+),敏捷 ($200M+).这些公司是前收入或早期收入,估值是由一般劳动力巨大的TAM驱动的.
- **操纵的基础模型:**物理智能 ($400M+),技能 ($300M+),共变 ($222M).投注是机器人AI遵循LLM扩展的玩法 - - 更多数据和计算产生更有能力的模型.
- 硬件启用者: 单位 (约合150亿美元以上),Flexiv (约合100亿美元以上).这些公司今天向真正的客户出售真正的硬件,并随着人工智能繁荣增长的收入.
对于进入这个领域的初创公司,我们建议专注于垂直应用 (特定行业的特定任务) 而不是竞争于一般用途的AI或硬件.
对于田野的成功意味着什么
- **如果Pi/Covariant基础模型成功:**机器人编程成为快速工程.瓶完全转向硬件质量和数据收集 - - 正是RCSV数据服务提供的.
- **如果Figure/Apptronik在规模上部署:**人体进入物流和制造业,从而产生了对培训数据,远程操作系统和维护基础设施的巨大需求.
- **如果Unitree的G1扩大:**大学实验室获得负担得起的人类型平台,加速研究.RCSV已经支持G1在我们的 租计划 和 数据平台.
- **如果开源 (Octo,Genesis,LeRobot) 获胜:**进入障碍下降,更多的初创公司进入,以及对质量数据和可靠硬件的需求增加.
