🤖 Understanding what exactly happens during your robot policy deployments is key to improving it. But interpreting raw rollouts is hard because they lack structure.
Introducing Behavior Graphs: turn many rollouts into an interpretable graph of what the policy is doing.
🧵👇
How can robots self-improve in the messy real world?
At #CoRL2026 with the first papers of my postdoc at @StanfordIPRL , led by the amazing students!
Would love to chat, DM open! I’ll be at:
Today (Mon), 5–6:30pm:
• MessyMem: Learning-from-doing memory for mobile manipulation — Poster #45, Lone Star 3F
• MessyNav: Zero-shot navigation via manipulation in messy environments — Poster #155, Griffin 2F
Nov 12 (Thu):
• Morning: co-organizing Learn@Deploy—how can robots keep learning after deployment?
https://t.co/YaBUJPGzi5
• 3:05–3:35pm: MessyMem in the SISREN spotlight session, Room 306
• 3:35–4pm: MessyMem & NavWorld posters at SISREN, Room 306
1/8
The real-world is messy. 🤖🧹
If there's no collision-free path, a robot shouldn't just stop, it should move objects out of the way. But which object can it safely move?
Introducing MessyNav: Zero-Shot Navigation via Manipulation✨ Accepted at #CoRL2026 🚀
🧵⬇️
7/8
🧐 Takeaways
🔸 MessyNav can be deployed zero shot in real-world settings.
🔸 Explicit affordance reasoning is key towards interactive navigation.
In future work, we will look at affordance conditioned skills and memory for long horizon navigation.
Check out MessyMem: https://t.co/UOXgqhExyQ
3/8
🤔 Why not let the planner figure it all out end-to-end?
Because affordance is contextual, not categorical:
🧳 A suitcase is pushable unless it has a "fragile" sticker on it.
🧱 Toy blocks are pickable unless they have been assembled into a child's castle.
Neither task planners nor category-level labels can identify the difference.