谁是藏镜人 (@VedaAI00) profile picture

探索科技与AI

Moon
Joined September 2024
Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)Photo by 谁是藏镜人 (@VedaAI00)
别把皮查伊发的「鹈鹕骑车」当普通玩具。 这次质变是把生成逻辑换成了「多 Agent 视觉自审闭环 (Critic Loop)」: Builder 在 Blender 建模与 Three.js 拼装,Critic 截渲染帧反查拓扑缺陷。代码写完不算完,直到视口真值自洽才通过编译。
1
2
377
过去 3D 生成卡在「长程幻觉」:大模型代码语法全对,渲染管线里全是破面与 UV 拉伸。 Antigravity 的 /swarm 让 Agent 形成视觉博弈,用渲染真值倒逼修复。这种自审闭环会先革掉谁? 传送门:teapawt.org/sundarpichai/status/21…
Carbon made a pelican demo so you don't have to 😅 fun test of long-horizon coding and the new /swarm feature in Google @Antigravity. Builder and critic subagents ran iterative loops across Blender and Three.js - checking their own renders to build a playable 3D game of a pelican biking around our Mountain View campus.
121
刚打开源码,不得不承认单文件 3D 渲染已被大模型玩出了花。 Gemini 4 Carbon 单 HTML 写出 3D 沿海赛车: 1. 零依赖管线:原生 Three.js 跑满 60 帧,内置物理碰撞与漂移粒子; 2. 算法声浪:Web Audio 合成 9250 转 V12 轰鸣与隧道混响,零外部音频。
1
2
641
大众以为大模型只会写 CRUD。当单 Prompt 能闭环物理引擎、光影渲染与程序化声效,浏览器的运行时潜能才真正被榨干。 端侧免安装的即时 3D 体验,正在重构轻量交互资产的生产链路。 原推:teapawt.org/StatsWire/status/21202…
Gemini 4 Carbon: Built a playable 3D V12 hypercar coastal racing game in a single HTML file with Three.js Sound ON: 9,250 RPM V12 wail, dual-clutch shift barks, stone sea-tunnel reverb, smoky hairpin drifts & 340 km/h viaduct overtakes at 60fps.
1
91
汽车拆解泰斗 Sandy Munro 团队,把宇树 G1 人形机器人全拆了。 Munro 硬件逆向实录: 1. 全铝 CNC:弃用开模压铸,关节件纯铝 CNC 掏空,用机加成本换极速迭代; 2. 驱控架构:弃用昂贵谐波,自研 15:1 双级行星齿轮,兼顾力矩透明度与抗冲击。
1
302
传统人形贵在开模与定制传感。G1 用低减速比行星齿轮实现高反驱,靠电机电流反推关节扭矩,砍掉昂贵传感器。 这不是象牙塔产物,而是硬核制造哲学:用供应链工程把科研机做成消费品。 原推:teapawt.org/live_munro/status/2119…
Understanding how the Unitree G1 is designed, engineered, and manufactured starts with examining how the product is built. At Munro & Associates, we recently completed a full physical teardown and engineering assessment of the Unitree G1, examining its architecture, actuators, electronics, materials, and structural systems. By tracing the product down to its individual components and manufacturing processes, we built a clearer picture of how the G1 is designed, assembled, and manufactured. The assessment connects physical teardown findings to engineering and manufacturing analysis, including a detailed production-cost model that accounts for components, materials, processing, and assembly. This provides insight into the decisions that shape product cost, complexity, and manufacturability. For organizations developing humanoid robots, evaluating competitive platforms, or looking to improve product cost and manufacturability, Munro provides the engineering analysis to support those decisions. Interested in the full Unitree G1 assessment? Email [email protected] to learn more about the findings and deliverables.
86
高精空间测绘昂贵,过去依赖专业激光雷达与多天扫描。 开发者仅用一台全景相机录制 5 分钟,本地跑通拉什莫尔山 3DGS 空间捕获: 1. 全景反求:从 360 视频连续抽取视差解算相机位姿; 2. 强光抗噪:在烈日与人群穿梭下,还原巨石雕像毫米级空间几何。
1
1
11
1,078
过去重建国家地标需数周航测。当 3D 高斯泼溅把普通全景视频转化为辐射场真值,空间计算已被消费硬件干平。 数字孪生不再是测绘巨头专利,而是现场漫步 5 分钟即可交付的数字资产。 原推:teapawt.org/AKurian001/status/2119…
Real-time walkaround inside a 3D Gaussian splat Spatial Capture of Mount Rushmore National Memorial. Processed and rendered locally in AirVis Studio from a single 5min 360 Video in Insta 360 X5. Sun was too bright and place was really crowded, lot of shadows. View in 3D : airvis.com/view/EmaAgvjvfOji… superspl.at/scene/2b267514
92
地面还在卷变压器和电网时,SpaceX 已经把算力能源搬上了太空。 锁定德国 SINGULUS 超 100GW 异质结产线,专供 STARMIND 轨道 AI 卫星: 1. 重工业自造:真空镀膜全套设备运抵德州 Bastrop; 2. 空间超算闭环:甩开地面散热与电网,太空直采光能驱动集群。
1
256
地面数据中心死穴在电网接入周期与巨量散热。轨道算力把冷源交给宇宙真空辐射,能源交给 24 小时光照。 若轨道算力成本跨过临界点,先被冲击的是陆基 IDC 电力,还是跨洋光缆? 原推:teapawt.org/ns123abc/status/212025…
🚨 BREAKING: SpaceX just locked up 100+ GW of solar cell manufacturing equipment from German toolmaker SINGULUS to build STARMIND AI satellites The first turnkey heterojunction line arrives in Bastrop, Texas in H1 2027, running wet chemistry, vacuum coating and thermal processing under one roof. Elon Musk says most of the mass of a Starmind satellite is solar panels, meaning the solar cell factory is the satellite factory. 100 GW is enough solar for 400,000 Starmind sats “Time to make the mind of a sentient sun.”
57
谁说移动端跨平台做不出主机级粒子特效? 开发者用 React Native + Expo WebGPU,把任意原生 UI 视图直接风化碎裂成数十万颗物理微尘。 当 WebGPU 计算着色器下沉到移动端,前端与图形学的边界正在彻底抹平。 原项目:teapawt.org/sumitXbravo/status/211…
Dust Disintegration 🫰 Wrap any view and it crumbles into dust. ✦ Built in React Native with @expo · WebGPU
119
别迷信单体大模型包揽一切,这台微型机器人的解法极具极客美感。 7 种运动技能拆成 7 个独立神经网络,毫秒级无缝热切换: 1. 技能解耦:平地行走、跌倒爬起各由独立策略接管,杜绝遗忘; 2. 极简管线:仿真训练 ➔ 轻量蒸馏 ➔ 仲裁器动态接管,端侧极省。
2
161
单体大策略总想兼顾所有工况,边缘状态极易崩溃。把动作原子化,用离散仲裁器热插拔,反而跑出了近乎生物本能的敏捷。 物理 AI 的落地,从来不是盲目堆参数,而是工程架构在端侧的极致克制。 原推:teapawt.org/0xKrosh/status/2119135…
This mini robot has 7 skills and 7 separate neural networks that swap on the fly. Walking, getting up after a fall, tricks: every skill is its own policy. Any of them can take over the robot at any moment. The development loop has four steps: train in simulation, deploy to the robot, refine the simulation from real-world failures, publish. A finished policy can be published for other owners to download, the way Hugging Face made sharing trained models routine. Sitting and standing can be triggered with a laser dot, without a single line of code. One honest note: camera resolution, LiDAR range and radio versions are, according to Pollen, not final yet, so the spec sheet may change.
55
当行业还在为几十万的科研人形买单时,CMU 团队把成本打到了 1500 美元。 开源双足机器人 BRIDGE 亮出硬解: 1. 构型-控制协同:废除硬件做完再调参的旧路,本体结构与控制策略联合优化; 2. 21 自由度:开源 CAD 与权重,实机完成翻滚跳跃高动态姿态。
1
3
1
15
1,246
传统双足昂贵,往往是用过剩电机扭矩去填补动力学缺陷。BRIDGE 用数据优化质量分布,从根上挤干硬件水分。 当千元级开源机体能跑通全身高动态,具身硬件的泡沫才真正被戳破。 原推:teapawt.org/techniahqrobot/status/…
This might be one of the most impressive low-cost humanoid robots I've seen so far Meet BRIDGE an open-source humanoid developed by researchers from Carnegie Mellon University and their collaborators. The robot costs around $1,500 to build yet it can perform backflips dance, balance on one leg and reproduce complex human movements.
77
机械臂开放世界操作最冗余的,是把控制、仿真与进度监控拆成三套独立架构。 斯坦福在 CoRL 推出 8B 具身模型 PsiBorg: 1. 统一自回归:动作策略、物理演化与任务进度统归为 Next-Token; 2. 零样本可控:告别复杂胶水,单模型输出完整操作概率分布。
1
7
533
传统管线用 ROS 拼接零散系统,时延与误差在闭环中成倍放大。PsiBorg 把物理推演直接做进 Token 转移概率。 当策略与仿真在自回归中合一,机器人控制正踩上当年大语言模型统一收敛之路。 原推:teapawt.org/AtlasKazemian/status/2…
Robot Capabilities as Probabilistic Inference over Next Token Prediction We are doing a demo at CoRL debuting PsiBorg: an 8B universally promptable robot model for open-world manipulation! One model works simultaneously as a policy, a simulator, and a progress tracker. It is trained from scratch, developed fully in-house at #Stanford, not a VLA or WAM, and (tries to) work with new objects and new goal images you throw at it! Come try it at #CoRL2026! 🕛 Mon Nov 9 · 12–6:30 PM CST 🕙 Tue Nov 10 · 10 AM–6:30 PM CST 📍 Demo #2 · Lone Star (3rd floor), JW Marriott Austin 🌐 neuroailab.github.io/psi-web…
1
75
人形机器人全身运动控制最头疼的,是换个机体就得从头调参。 前 OpenAI 科学家 Andrew Carr 联合 Menlo 开源 Cartwheel Control: 1. 硬件解耦:运动学与本体动力学分离,控制策略跨硬件通用; 2. 开源落地:首发开源宇树 G1 等三套全身动作追踪 Policy。
2
2
16
1,095
以往动捕重定向频频摔倒,死穴在各家电机响应和触地刚度各异。Cartwheel 先学物理自洽的平衡解,再映射到末端扭矩。 当控制底座通用化,人形机器人的护城河就不再是调参,而是动作数据规模。 原推:teapawt.org/andrew_n_carr/status/2…
We are pleased to announce Cartwheel Control! Our hardware-agnostic approach to training whole-body motor control policies from large-scale kinematic motion. To kick it off, we’re releasing three open motion-tracking policies for humanoid robots: Unitree G1 @menloresearch Azimov @RoboParty_DEV RP1 All free to download on Hugging Face.
94
谁说做 3D 激光雷达扫描必须上万元专业测绘仪? Laan Labs 用 50 美元意法半导体传感器(ST VL53L9CX),做出了极客空间扫描仪: 1. 极简硬件:开源 PCB 板载 dToF 阵列,USB 直连手机; 2. 软硬融合:借用 ARKit 解算空间位姿,实时融合成密集室内三维点云。
3
14
116
6,178
机械雷达贵在光机与电机,固态面阵 dToF 结合手机 VIO 视觉里程计,把最贵的位姿解算成本一把干平。 当空间测绘硬件降到两顿饭钱,最先普及的是家装数字孪生,还是个人具身机体环境感知? 原推:teapawt.org/laanlabs/status/212020…
Testing 3d scanning with this ~$50 lidar sensor (ST VL53L9CX) attached to an iPhone. Works pretty well for quick indoor scans using just ARKit poses. PCB is open source from RemiFM: github.com/RemiFM/USB-LiDAR-…
2
341