PhD student @Harvard Reinforcement learning & Robotics Ex @GoogleDeepMind @Tsinghua_Uni
Cambridge, MA
Joined February 2025
- Tweets12
- Following38
- Followers33
- Likes12
Kimi K3.1 pro push the frontier of open-sourced models in robotics tasks!
🚀 Open-weight models are catching up fast in robotics!
Kimi K3.1 Pro just landed at #7 overall and #1 among open-weight models we tested on RLE-Bench!
This is another exciting milestone for open-weight AI. The gap between open and proprietary frontier models is narrowing on complex robotics tasks that require physical reasoning and experimentation. Exciting to see more capable agents becoming accessible to the broader robotics research community.
🔗 Full leaderboard: rle-bench.github.io
#RLEBench #Robotics #OpenSource #Kimi #EmbodiedAI
1
43
Gemini 4 Argon performs strong on our RLE-Bench! 🚀
Check out full results on our website rle-bench.github.io
🚀 Gemini 4 Argon takes #1 on 4 robotics tasks in RLE-Bench!
The latest Gemini 4 Argon results are in: 72.4 RLE Index, ranking #2 overall, just behind GPT-6 Astra (73.4).
More impressively, Gemini 4 Argon takes the top spot on 4 out of 9 tasks:
* Harness Engineering
* Embodied Reasoning
* Whole-Body Motion Tracking
* NanoVLA Recipe
It’s exciting to see frontier AI labs continuing to invest in robotics capabilities, with each new generation pushing the boundaries of what general-purpose AI agents can accomplish in embodied and physical-world tasks.
The competition is getting closer, and the progress is accelerating.
Looking forward to the Carbon! 👀
Explore the full leaderboard: rle-bench.github.io
#Robotics #EmbodiedAI #RLEBench #Gemini4
75
Nice work! RSI that achieves 100% success rate in RoboCasa 🚀🚀🚀
Check the website for details: stacx-lab.github.io/physical…
Coding agents can solve all RoboCasa tasks from scratch, by self-evolving its skills and tools from its own trials.
Our system grew its library from empty to 6 canonical skills and 70 functions in seven rounds, and solved all 365 tasks.
See details: stacx-lab.github.io/physical…
8
393
Excited to share what we have been building for the last few months!
Kinex defines essential interfaces connecting agents to the real world. We have demos on different embodiments including fixed, wheeled and legged manipulators. It autonomously onboards itself to the robotics by writing device drivers.
Check out more zero-shot real-world demos on our website stacx-lab.github.io/kinex/
Introducing Kinex, the harness that connects coding agents to any physical robot, unlocking scalable learning in the physical world.
Tell Kinex about your robot; it then carries out tasks and learns from every run. We demonstrate Kinex across diverse real-world applications in a zero-shot setting, from bimanual manipulation to quadruped robots.
🧵1/5
4
63,957
Two updates from RLE-Bench 🚀
Glad to see the frontier models are continuously improving on the robotics tasks. Check our technical report in the comments to see the details!
Two updates on RLE-Bench! 🤖
📈 New results for GPT-6.1-Sol, Opus 5.5, Fable 5.1, GPT-6-Luna, and Grok 4.6! Opus 5.5 shows substantial gains in interactive control—even outperforming Fable 5.1. Exciting to see frontier models making progress on robotics tasks!
📑 Our technical report is out! We dive into the design principles behind RLE-Bench, break down each task, and share observations of how agents tackle them. Plenty of interesting agent behaviors to explore!
Check out the latest results and read the report 👇
Thanks Kairun! Nice work on the summary!
1
169
A few months ago we saw Fable 5 and Kimi k3, we started exploring a direction beyond VLAs and world models: what if general-purpose coding agents could actually do robot learning engineering?
Now GPT-6 Astra is making that question feel a lot less hypothetical.
GPT-6 Astra is impressive, but agentic robotics ≠ just control.
Physical agents should control robots, learn new skills, and design their own hardware.
So we introduce RLE-Bench: 48 everyday robotics engineering tasks spanning closed-loop control, policy learning, perception, and mechanical design.
4
298