Chunru Lin retweeted
I'm at #CoRL2026! Excited to present our work, HARP-Bench, tomorrow!
We introduce a benchmark and data generation framework for evaluating and improving robot safety in human-present manipulation.
📍 Nov 10, 5:00–6:30 PM | Poster #89, Lone Star (3F)
🌐 harp-bench.vercel.app/
💻 github.com/UMass-Embodied-AG…
🎉Excited to share that RoboWits has been accepted to #NeurIPS2026!
Robots in the real world keep running into the unexpected. A task solution may not match the task instruction; can robot policies nowadays reason, adapt, and use tools creatively when that happens? 🧵
Update: We just benchmarked Fable 5.2 and Astra! 📊
The TL;DR: Fable 5.2 > Astra >> VLAs.
Check out the updated table and detailed results in the repo👇
github.com/UMass-Embodied-AG…
18
⚒️It's fully open and easy to plug in:
• LeRobot-compatible, GPU-accelerated
• Demo dataset on 🤗
• RoboWits-10 split for standard comparison
📰 arxiv.org/abs/2605.30326
💻 github.com/UMass-Embodied-AG…
🤗 huggingface.co/datasets/XHRl…
1
1
36
Benchmark your policies or agentic robotic frameworks on RoboWits and share your findings! We'd love to see what closes the gap! 🚀
Huge thanks to my amazing co-authors @icefox1104 @FenghaoYu @ZhehuanChen @cocosci_lab @YejinChoinka @davheld @gan_chuang
1
51
Check out our latest work on VLA test-time adaptation! (P.S. I'm at IROS 2026 in Pittsburgh and happy to chat! Just DM me~)
Most VLAs are memoryless: when a robot's hardware drifts, the policy repeats the same failure every trial.
What if it could learn from its own failed rollouts and fix itself, without tedious recalibration?
Introducing a post-training recipe to remedy this: Self-Adaptive VLA 🧵
35




