与「PhysicalAI」相关的搜索结果

看机器人踢球，感觉是挺简单。 给机器人设几个动作，跑过去、抬脚、射门，好像就能完成。 现实里的机器人面对的是各种变化：球的位置、身体平衡、环境干扰，每一步都不是提前写几个指令就能解决的。 这也是为什么我最近比较关注 @axisrobotics。 Axis 做的事情，本质上是在解决 Physical AI（实体 AI）怎么让机器人真正学会行动。 最近 Axis 相关研究登上 Science Robotics，重点就是让人形机器人通过视觉理解环境，再学习复杂动作。 以前机器人很多动作靠人工设计。 但 Axis 关注的是另一种方式： 让机器人在模拟环境里不断训练，学习怎么控制自己的身体，然后把这些能力放到真实机器人里验证。 比如足球这种场景，需要机器人同时处理： 看到球在哪里； 判断下一步怎么移动； 控制身体保持平衡； 完成传球、射门等动作。 这些东西靠人工一个个写，效率太低。 所以现在 Physical AI 的竞争，本质上就是谁能让机器人获得更好的学习能力。 Axis 最近一直在推进这条路线： 从 Axis V2，到数据采集系统，再到社区贡献的数据体系，核心都是围绕一个问题： 怎么让机器人获得更多、更真实的数据，然后不断提升表现。 当然，现在机器人行业还在早期。 成本、稳定性、应用场景，都需要时间验证。 但我觉得 Axis 值得关注的地方，是它没有只停留在“机器人展示”这一步，而是在做机器人背后的训练和数据基础。 未来机器人能不能真正走进现实，关键不只是造一个机器人出来。 更重要的是，机器人能不能持续学习。 这也是 Axis 现在在押注的方向。 @axisrobotics #AxisRobotics# #PhysicalAI#

最近几天刷到一堆人在玩同一个东西，机械臂 @axisrobotics 不是真去工厂里搬箱子那种，是坐在电脑前，用浏览器远程操控一只虚拟机械臂，去抓豆腐、放盘子、开垃圾桶盖……听起来挺简单，真上手才发现有多折磨人。方向稍微偏一点就抓空，夹爪扭得像麻花，明明看着已经对准了，最后还是差那么一毫米。好几个手残党朋友跟我吐槽：玩了一个小时，一个任务都没过。 但奇怪的是，越玩越上瘾。 Axis 干的事其实挺反直觉的。很多人以为机器人训练就是拼命让人多操作、多教，操作越多数据越好。结果他们发现，人类操作里有大量“低信息量”的动作——机械臂空着穿越空间、停顿、反复微调、绕远路，这些动作成本很高，教给模型的东西却很少。V2 直接把这些砍掉了。你可以拖动末端执行器，或者直接点选目标物体，让它自动挪到预抓取位置。真正留给人操作的，只剩抓取选择和夹爪闭合前后那几秒的精细调整。 更狠的是纠错数据的处理方式。模型自己跑任务，快失败的时候人接管把它拉回来。以前大家可能会把整段人类操作全录下来喂给模型，结果模型反而变笨了。后来他们只保留那些真正把机器人从失败状态救回来的短动作，平均只有 0.8 秒。过滤完之后，成功率明显上去了。 这套逻辑有点像教小孩骑自行车。你不会全程扶着他走，也不会把每一次歪歪扭扭都当成标准示范。你只在他快要倒的时候伸手扶一下，让他自己学会从失衡里恢复平衡。Axis 现在做的，就是这个。 普通人不用买机器人、不用搞 GPU，打开浏览器就能贡献数据。任务从简单的抓取放置开始，后面会慢慢往长时程、多步骤、需要容错和协调的方向走。现在看起来还像抓娃娃机，几个月后可能就会变成真正的“高级钳工”考核了。 数据会上链签名验证，贡献是实打实的。再加上和 Kaito 的创作者计划，写内容、拉人参与，都能计入最终的奖励权重。奖励池是 $AXIS 总供应量的 0.25%，TGE 一次性解锁。 说白了，这不是单纯让大家玩游戏，也不是单纯让大家写推文。它在尝试把“机器人怎么在真实世界里犯错、怎么被纠正”这件事，拆成普通人也能参与的碎片。以前这些数据只能在实验室里一点点攒，现在变成了浏览器里的日常任务。 我自己也试了几天。一开始真的很挫败，后来慢慢摸出点门道：先让机械臂到物体正上方再下降，靠近后再微调，别一上来就大幅晃。完成任务后记得去看验证状态并签名，不然轨迹不算数。 玩着玩着突然觉得挺有意思的。机器人想走进现实世界，缺的从来不是更多算力，而是足够多、足够真实的“怎么从错误里爬起来”的数据。现在这些数据，有一部分是由坐在电脑前的普通人一点点攒出来的。 如果你也想体验一下当“远程钳工”的感觉，可以去试试。难是真的难，但偶尔一次抓成功的瞬间，还挺解气的。 快来试试看吧❤： @axisrobotics @KaitoAI #AxisRobotics# #PhysicalAI# #KaitoAi#

The future of humanoids isn't just mechanical—it's neural. A humanoid robot is an embodied AI system. Motors, sensors, and actuators provide the body, but neural intelligence provides the mind. Neural models power vision, speech, language, navigation, manipulation, planning, memory, decision-making, and continuous learning. As robotics evolves, nearly every cognitive subsystem is becoming neural-first. That's why represents more than a niche—it's a foundational concept for the next generation of intelligent robotics. #HumanoidRobots# #EmbodiedAI# #PhysicalAI# #NeuralNetworks# #ArtificialIntelligence# #Robotics# #MachineLearning# #DeepLearning# #RobotLearning# #ReinforcementLearning# #ComputerVision# #GenerativeAI# #AIResearch# #Automation# #FutureOfAI#

🚀 Today we're launching agentic triage in @roboto_ai. Roboto Agents can now review every robot log the moment it arrives and open labeled issues in #Jira#, #Linear# or #GitLab# with supporting evidence attached. It’s the first line of defense that every #robotics# team needs. Failure modes and anomalies are detected automatically. The runs that warrant escalation get surfaced. You can spend more time improving your system instead of debugging the same issue for the fifth time. Robotics teams often underestimate how high the support burden will be once they start putting robots into the field. A robot that's 99% reliable sounds great until you deploy 100 of them. Suddenly you're dealing with failures every day. The statistics catch up quickly. Tomorrow’s issues are already sitting in today’s logs - but nobody has time to review them. So you wait for a customer to complain instead. That’s a huge business risk. Agentic triage gives you the tools to find issues before your customers do. We’ve been testing this with select engineering and field support teams collecting #ROS# bags, #MCAP# and #PX4# ULG files over the past few months. They think it rocks, and they’ve discovered and fixed some surprising issues before anyone noticed. Today is day 2 of #AgentWeek#, and it's live now. If you want to try agentic triage on your own data, reach out or check out our blog post below. #PhysicalAI# #Triage# #Agents# @OpenRoboticsOrg @PX4Autopilot @NVIDIARobotics

WAIC 2026 — Come meet our busy bee Oli at Hall H3, Booth C-216. Our Full-Size Humanoid Robot Oli has been keeping the crowd engaged. Autonomously walking up slopes and stairs, and delivering a live demonstration of full-body teleoperation powered by LimX's whole-body motion foundation model. Powered by COSA, LimX's humanoid brain system, Oli can chat with you in multiple languages, solve riddles, and respond to voice commands, such as breaking into a dance whenever you ask. Come put Oli to the test. #WAIC2026# #WAIC# #PhysicalAI# #Humanoids# #Robotics# #AI# #LimXDynamics#

Your best Gaussian Splat might already be sitting in your Scaniverse account. The #SplatYourWorld# Challenge is open through July 31. Submit your best capture for a chance to win a share of $4,000 USD in prizes. We're looking for standout 3D reconstructions of: 🏭 Industrial environments 🌲 Natural landscapes 🏙️ Urban infrastructure No need to create something new—just submit your best Gaussian Splat and put the prize money toward new gear, more captures, or whatever powers your workflow. 📅 Deadline: July 31, 11:59 PM PT Learn more: #NianticSpatial# #3DReconstruction# #DigitalTwins# #PhysicalAI# #GeospatialAI# #AI# #3DGS#

NVIDIA just dropped Cosmos 3. The first open foundation model that reasons about the physical world. And tells a robot how to move. We've had GPT for words. This is GPT for atoms. 🤖 #PhysicalAI# #FrontierTech# #Robotics# #AI#

Welcome to The Robot Almanac, Morgan Stanley’s new series exploring the rise of robotics and physical AI through 2050. Watch the introduction now and subscribe for future episodes.

BIG ANNOUNCEMENT FROM HUGGING FACE TODAY: We're unveiling Microduck 🐥🤖 It's a tiny $399 open-source robot you can teach new tricks with reinforcement learning. It can walk, pick things up, get back up when it falls, and even roller-skate. Welcome to the era of open-source affordable robots to democratize physical AI and world models! 🤗🤗🤗

From Digital Twins to Data Engine: Cutting the Real-Data Burden with Sim-Powered Robot Learning Teaching a robot a new task could take hundreds of teleoperated demonstrations. For foundation models, adapting to an entirely new robot can cost orders of magnitude more — dedicated hardware, trained operators, months of engineering. On @boosterobotics' dual-arm robot, we studied this at two levels: ✱ Specialist: Can task-aligned simulation mixed with a small set of real demonstrations reduce the real-data burden? ✅ Yes. With only 10 real demos, the policy made no contact at all in physical rollouts (0/20). Adding 50 simulated trajectories brought contact to 17/20. ✱ Foundation: Can data accumulated across tasks build a reusable starting point (a Booster-specific model prior)? ✅ Yes. After full-parameter continued pretraining, a model adapted with just 30 demonstrations per task beat the original given twice as many: 14/16 vs 10/16 in simulated evaluation. Before any task-specific adaptation, in zero-shot simulation, it was already roughly 3× closer to the target (17.27 cm → 5.78 cm). This work runs on Axis Suite, our Physical AI solution across different robot embodiments. Distributed contributors generate task-aligned sim data on Axis Hub at scale, reducing real-data needs for specialist adaptation while powering cross-embodiment generalist training. Read the full blog: