Peng Xu (@sippeyxp) header
Peng Xu (@sippeyxp) profile picture

Working on #Gemini Robotics model in @GoogleDeepMind; lead Gemini Robotics Trusted Tester program; making AI-powered robotics accessible to all.

Joined February 2010
Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)Photo by Peng Xu (@sippeyxp)
♊+💡+🖨️ = Happy Halloween🎃🦿
While most jack-o-lanterns just sit on porches, the @GoogleDeepMind team decided to let AI agents design, simulate, and bring one to life. 🎃🤖 We built a physical, dancing 12 degrees of freedom robot directly from a single plain-English prompt, with no manual CAD required!
3
291
This design is so practical and safe comparing to 50kg robots. Kinda want one for the kids room.
Stringman MCP and Opus 5.5 clean a playroom Some help was given when it asked for several large items to be set aside but it did OK. 129 minutes total. $795 at neufangled.com/
4
880
Meet Sprout, a semi-autonomous plant robot powered by #Gemini and build with @antigravity. Watch as it navigates the office in search of sunlight, pings for help, and uses Gemini Live and Robotics ER 2.0 for vision to care for fellow office plants. youtu.be/ID6vEy94j7c?si=rLlv…
1
2
2
26
6,346
The competition is on!
Gemini 4 Argon achieves SOTA on real robot capabilities, tied with Opus 5.5 and above GPT-6 Astra. 🧵
2
1
39
1,650
🔥Exciting times! Apollo having #tailgating party fun w/ #Argon.🏈🌽 It watched human play, real2sim, developed the motion in the sim and iterated action on real robot to score🎯! #Robotics team is super hyped by #Argon's capabilities and you should try it on your robots too!
Replying to @Google
In robotics, Gemini 4 Argon is speeding up real-world physical learning. We let Argon watch a video of a human playing cornhole. It mapped the environment, calculated whole-body joint movements, and autonomously built the simulation code to practice. When transferred to Apptronik’s Apollo humanoid, the robot sank seven straight shots in a real-world trial.
7
600
🗓️9 weeks ago, I wrote about future physical AI ecosystem needs benchmark agencies📈. Here it comes! Really like @Nicolas_Keller and @josermedinahdez 's work: 1) Manufacturing themed tasks - rare. 2) Show data performance trade-offs. 3) Controlled eval to reduce variance.
The Era of Evals is coming to AI robotics. Today we’re launching Reality Check: our leaderboard & first public robot manipulation benchmark. 14,400 real-world rollouts. Four models. Multiple tasks & data regimes. Object placements. Confidence intervals. FR3 Duo stations. 1/10
1
6
1
20
9,783
😱 /goal maximize 🖇️ production -> 💥
hilarious to watch how Opus 5.5 designs a 3d asset, prints it, and controls a robo arm to take it out of the printer
1
9
1,520
I saved all my chat since I started using WeChat. Time to try it out!
一直在等微信开放接口,让我可以处理积累下来的上百 G 聊天数据。试过 Github 上的几个项目都不太成功。 万万没想到,最后这个接口来是被腾讯开放了 :) 感谢 HY4 preview 非常到位 :D 模型太强了!
3
623
I was here in all-in summit. Life is a reality show😂
Holy crap POTUS just phoned in @JensenHuang live on stage at All In Summit We will not lose the ai race! And whatever Dario said this weekend won’t stop our progress This made my morning! $NVDA
1
2
1,327
This is huge! 🤯 Reliance on real robot data and cross embodiment are big issues of robotics action models! Super excited to see OM1 solving both of them! The motion is also fast and looks smart! Great work @zipengfu @chenwang_j!
Introducing OM-1, our first robot foundation model, zero-shot generalizing to any robot: table-top arms, industrial arms and humanoids. - learned directly from human manipulation data - no teleop/robot data - close to human-level dexterity and efficiency - multi-robot collab
11
23
149
701,002
Handling noises well. Great! I have a challenge: take the cute reachy bot into a bar, and maintain a conversation with it. This is a real challenge every time I bring #pupper into to a party. @OpenAIDevs do you want to take this challenge?
GPT-Live-1 is now available in the API. Bring ChatGPT’s natural back-and-forth to your app, with voice agents that listen while they speak and work with the models and harness you choose.
1
1
5
1,037
🏷️50% price reduction of #Gemini #Robotics ER 2 💥 We have seen a lot of love of #Gemini #Robotics ER 2. One complaint we heard is that the model is more expensive than the previous ER versions. Here you go, FIXED 🛠️! ai.google.dev/gemini-api/doc…
🆕Introducing #GeminiRobotics ER 2 @google ER 2 packs a stronger temporal understanding: an important aspect of embodied reasoning. It pinpoints key moments in video, understands event progression, orchestrates VLA and robot APIs, etc. More👇 blog.google/innovation-and-a…
4
1
15
1,079
Peng Xu retweeted
GPT-6 Astra doing PCB layout in KiCad 🤯🤯
92
457
151
5,010
966,686
"Your mind is clear!" So much childhood memory! How about also "Theme Hospital" and "Caeser III"? @davidkal88
Wow “Make Red Alert 2 (YR) compile and run natively on iOS & macOS.” Codex (GPT-5.6 Sol) worked for 26 days, analyzed the EXE and rebuilt the entire game in 624K lines of C++ This is by far the hardest task I’ve ever given Codex as the game's code was never released and is believed to be lost. >
Readers added context they thought people might want to know
There is already an open source effort to build this and the posts claim that it was decompiled and built without assistance is false as the information to unpack the assets and game mechanics and rendering is already in the training corpus github.com/hel1o-wor1d/Op…
1
530
Unlike last time, Gemini 3.8 Flash is at true pareto frontier of price x performance space. Fake it until you make it in foundation models 😁 It also sits nicely at the sweet corner spot - for massive adoption.🚀
3.8 Flash = great quality @ great price
1
7
930
Robotics has been bombarded by model announcement in August. Beyond in-context learning, the mission statement at the end of video depicted an exciting future. "Every small business, every grocery store, every hospital, every organization can have the same power of robots."
Introducing S1, our new foundation model that learns from one example. It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning. Watch S1 operate in real-time via in-context learning:
3
1
1
12
2,004
Gemini 3.7 outputs joint actions directly to complete a robot task! Great work from @chooi_jeq. This and recent work such as CaP-X and ENPIRE, etc suggests coding for robotics is making a strong return. Combine coding with VLA / WAM is a good topic to think about.
Gemini 3.7 Flash just saturated one of our physical tool-use benchmarks at 92%. Gemini 3.6 Flash, released three weeks earlier, scored only 32%. This marks a step change in the robotics capabilities of LLMs. 🧵
7
12
56
7,773
Supermarket robot hosting a coconut water taste stand. Robot from DexTeleop dexteleop.com. Fully teleoperared. Operator can hear but does not speak. Deployed last month, it is not longer making a scene Highlight 👧🏻: Robots are replacing human. 🤖: 😱[dropped cup]
Wow my hometown! I am in Taiyuan visiting my family, gonna check it out tomorrow😃
3
21
2,473
Wow my hometown! I am in Taiyuan visiting my family, gonna check it out tomorrow😃
Two wheeled humanoid robots are being deployed at JD's 7FRESH supermarket in Taiyuan… They are assisting with restocking, shelving, and providing product tasting and guidance. (VR TeleOP) From Beijing to Tianjin and Taiyuan…JD is introducing humanoid robot employees to its retail supermarkets in large numbers to enhance the shopping experience and alleviate peak customer traffic. But now,more importantly,TeleOP data These humanoid robots are from Beijing-based DEXTELEOP's TA2… who have also entered into a strategic partnership with JD. They are being deployed through JD's nationwide retail channels(20k). Placing data collection in real-world operational scenarios and achieving commercial success is a significant achievement. The physical world is the largest data collection center for humanoid robots.
1
1
25
5,574