Skip to content

Day 1: Mon, Nov 9, 2026

Sessions attended

Papers covered

Paper Group Headline Status Links Weights
Being-H0.5 BeingBeyond 35k h across 30 embodiments with a unified action space; 98.9% LIBERO abstract arXiv · page ?
DreamZero NVIDIA 14B video model as policy; over 2× generalization vs VLAs at 7 Hz summarized arXiv · page · code ✅
LVP MIT, UC Berkeley, Harvard 14B video planner for zero-shot robot control; 59.3% task completion on 100 third-party tasks vs Wan 2.1, Hunyuan, Cosmos talk-only page ?
MemoryVAM UCLA, UCSD, Utah, Nirvana Robotics LIBERO-Mem from 5% to 42.5%; 75–80% on real memory tasks summarized arXiv · page ?
Mem-World Future actions pick which history a multi-view world model retrieves, via a wrist-view 4D surfel memory talk-only ?
MolmoBot Ai2 Sim-only training; 79.2% real pick-and-place vs 39.2% for π0.5 abstract arXiv · page ?
π0.7 Physical Intelligence Metadata turns low-quality data from harmful to helpful; zero-shot UR5e shirt folding at teleoperator level summarized arXiv · page —
RoboTTT NVIDIA GEAR, Stanford, UT Austin 8K-step context at constant per-step cost; 79% vs 42% task completion summarized arXiv · page ?
TEMPO UC Irvine Motion and command history for +2M parameters; 47.5% to 76.0% average success summarized arXiv · page · code —

Day 1 themes

Cross-day versions: world models, memory, human data, Systems 0/1/2.

  1. Scaling along new axes: Data diversity with metadata (π0.7), human video (Being-H0.5, DreamZero), context length (RoboTTT), and simulation scale (MolmoBot).
  2. Cross-embodiment transfer is now the default goal: Every talk claimed some form of transfer to new robots or zero-shot deployment.
  3. Video and world models are moving into the policy: DreamZero predicts video and actions together, π0.7 uses a world model for subgoal images, and MemoryVAM adds memory to a video action model.
  4. Single-frame policies are not enough: RoboTTT (8K-step context), MemoryVAM (episodic memory), TEMPO (motion and action history), and π0.7 (MEM video history) all add temporal context. They differ in cost. TEMPO and RoboTTT keep the cost of each step constant, and MemoryVAM's cost grows with episode length.

Live notes

MARS Lab: Organizing Physical AI Across Timescales

About 10:00 am, from slide photos. Full notes are in themes/system-0-1-2.md.

Systems 2, 1, and 0 at their clocks, with intent going down and feedback mostly missing on the way up, and the two ways forward

System 0 / 1 / 2 stack

Being-H0.5

About 10:50 am, from slide photos. Full notes are in papers/being-h05.md.

Being-H0.5: the human hand and each robot write only their own slots in one shared action space

Oral Session 2, Talk 1: DreamZero

Slides photographed at 11:11 and 11:14 am. Full notes are in papers/dreamzero.md.

DreamZero closed-loop inference: inputs into a joint video-action DiT, future frames and an action chunk out, and real observations written back into the KV cache

Mem-World

Slide photographed at 11:25 am. Full notes are in papers/memworld.md.

Mem-World: future actions give a wrist pose that picks which past frames to retrieve from the surfel memory for the world model

Evening posters

Photos taken 5:30–5:53 pm.

LVP: image and instruction to a video plan, recovered hand motion, retargeting, and robot