Skip to content

CoRL 2026 Conference Digest

By Karthik Pythireddi

A digest of the talks, papers, and cross-cutting themes from the Conference on Robot Learning (CoRL 2026), Austin, Texas, November 9–12, 2026.

About the conference (from corl.org):

  • Dates: Main conference November 9–11, 2026, and workshops on November 12.
  • Venues: Hilton Austin for presentations and keynotes on Days 1 and 2. JW Marriott Austin for poster sessions, exhibits, lunch, and breaks. Bass Hall for the reception and the Day 3 keynote.
  • Program: 32 accepted workshops, and a demo track that requires a physical robot system.
  • Awards (reported in press coverage): Best Paper, SuperMimic (UC Berkeley). Best Student Paper, RoboTTT (NVIDIA GEAR, Stanford, UT Austin).
  • Scale (reported in press coverage; not published on corl.org): about 2,094 submissions and 687 accepted papers, an acceptance rate of about 33%.

CoRL 2026 at a glance

CoRL 2026 in four days: every talk in these notes, coloured by theme

Six themes ran through the week. Each page pulls together the talks across all four days, with tables and diagrams:

Theme The one-line version Page
World models Used three ways: as the policy (DreamZero, Dyna-2), as a simulator (Tesla, Hydra-0, Atlas), and as a planner (π0.7, LVP). A Day 4 workshop argued you may not need one. World models and world action models
Memory and long context Single-frame policies aren't enough. Memory ranged from history tokens and fast weights to text summaries and 3D object graphs. Memory and long context
Simulation and sim-to-real Isaac Lab 3.0 and Newton, Steel, and Arena for heterogeneous robots. Agents now build the sim from real video. Demo resets make RL practical for dexterity. Simulation and sim-to-real
Human data and scaling From 12 minutes (DreamZero) to 1M+ hours (Dyna-2) of human video. The hard part is mapping human motion onto robots. Human data and scaling
Deployment and reliability 99% per step finishes only 45% of a 79-step job. Industry talks were about recovery, fleet health, interventions, and diagnostic evaluation. Deployment and reliability
Systems 0 / 1 / 2 Most stacks split a slow planner from a fast policy. They differ in the interface between them. The feedback path back up is mostly missing. Systems 0 / 1 / 2

Days

Day Date Papers Talks
Day 1 Mon 2026-11-09 9 5
Day 2 Tue 2026-11-10 2 8
Day 3 Wed 2026-11-11 3 9
Day 4 Thu 2026-11-12 3 6

Talks and sessions

Every talk and session in these notes, including industry and tech sessions that have no paper file.

Day Talk
1 MARS Lab: Organizing Physical AI Across Timescales
1 Being-H0.5
1 Oral Session 2, Talk 1: DreamZero
1 Mem-World
1 Evening posters
2 Boston Dynamics: Single Action Model
2 Submerged manipulation: DAVATaR (UT Austin)
2 Torque adaptation: TAM (Ai2, KAIST)
2 Keynote: Russ Tedrake, Learning from the Factories
2 Meta Robotics Studio
2 AGIBOT: Scaling Robot Intelligence, Building a Physical AI Ecosystem
2 Tesla: Building Robot Foundation Models
2 Franka Robotics: Franka Research 3 and FCI
3 Tech Session: Isaac Lab 3.0
3 Skild AI: Steel: Simulation for Generalist Robot Learning
3 Video to Data: Agentic Native Pipeline
3 Hybrid Simulation: Cosmos Neural with Isaac Lab
3 FOVI-Dex: foveated perception for dexterous manipulation in simulation
3 NVIDIA Isaac Lab-Arena: one evaluation framework
3 RLWRLD: DexBench
3 Ken Goldberg: Agentic Robotics
3 Keynote: Fei-Fei Li, Atlas: A Unified World Model for Generation and Reconstruction
4 Workshop: Do Robotics Need World Models?
4 Talk: Long-Context, Long-Horizon, Compositional Generalization without World Models
4 Remote talk: VLAs, world action models, and two data premises
4 Workshop: Sim2Real2Field: A Model-Based Perspective
4 Michael Lutter (Boston Dynamics): Sim2Real2Field at Boston Dynamics: The Unglamorous Part of Scaling Humanoids
4 Dyna Robotics: Ingredients of Scaling Foundation Models in Deployments

Papers

Status: talk-only: 3 · abstract: 5 · summarized: 8

Paper Day Group Headline Status Links Weights
Being-H0.5 1 BeingBeyond 35k h across 30 embodiments with a unified action space; 98.9% LIBERO abstract arXiv · page ?
DreamZero 1 NVIDIA 14B video model as policy; over 2× generalization vs VLAs at 7 Hz summarized arXiv · page · code ✅
LVP 1 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 ?
Mem-World 1 Future actions pick which history a multi-view world model retrieves, via a wrist-view 4D surfel memory talk-only ?
MemoryVAM 1 UCLA, UCSD, Utah, Nirvana Robotics LIBERO-Mem from 5% to 42.5%; 75–80% on real memory tasks summarized arXiv · page ?
MolmoBot 1 Ai2 Sim-only training; 79.2% real pick-and-place vs 39.2% for π0.5 abstract arXiv · page ?
RoboTTT 1 NVIDIA GEAR, Stanford, UT Austin 8K-step context at constant per-step cost; 79% vs 42% task completion summarized arXiv · page ?
TEMPO 1 UC Irvine Motion and command history for +2M parameters; 47.5% to 76.0% average success summarized arXiv · page · code —
π0.7 1, 4 Physical Intelligence Metadata turns low-quality data from harmful to helpful; zero-shot UR5e shirt folding at teleoperator level summarized arXiv · page —
DAVATaR 2 UT Austin Fluidic tactile and immersion-depth sensing plus glare and zoom augmentation; 71% vs 41% mean success without tactile summarized page ?
TAM 2 Ai2, KAIST Sim-only residual torque layer below any policy; real Panda box pushing 47.6% to 76.2%, BC flip 50% to 72% summarized arXiv · page · code ✅
GaP 3 UC Berkeley, NVIDIA, Bosch Coding agents self-learn a graph of modular skills in a Newton Real2Sim, then export plain ROS 2 code; no frontier model runs on the robot abstract arXiv · page · code ?
Hydra-0 3 NVIDIA, Brown, Columbia, Harvard World model conditioned on action flow (robot actions as pixel motion); runs as a hybrid simulator, with physics for the robot and video for the world talk-only page ?
SuperMimic 3 UC Berkeley One learned world model + analytic gradients; beats SONIC's return in 3 h on 1 GPU vs 7 days on 128 GPUs (>1000× fewer GPU-hours) summarized arXiv · page ?
Dyna-2 4 Dyna Robotics World-action model pre-trained on 1M+ hours of human video only; zero-shot robot action MSE 0.180 to 0.117 from 1k to 1M hours abstract page ?
FARM 4 UC Berkeley, Stanford Real-time object-level relational 3D memory answers relational language queries; +164% Recall@5 over prior methods (unverified) abstract arXiv · page · code ?

Status levels: stub means a title only. talk-only means talk notes with no paper summary. abstract means a summary from the abstract and secondary sources. summarized means the paper's HTML or text was read. read-pdf means the full PDF was read, including the figures.

By tag

Tag Papers
cross-embodiment Being-H0.5, DreamZero, Hydra-0, LVP, π0.7, TAM
video-model DreamZero, Dyna-2, Hydra-0, LVP, MemoryVAM, Mem-World
simulation GaP, Hydra-0, MolmoBot, SuperMimic, TAM
vla Being-H0.5, MolmoBot, π0.7, RoboTTT, TEMPO
human-data Being-H0.5, DreamZero, Dyna-2, LVP
sim-to-real GaP, MolmoBot, SuperMimic, TAM
temporal-context MemoryVAM, π0.7, RoboTTT, TEMPO
world-action-model DreamZero, Dyna-2, Hydra-0, MemoryVAM
memory FARM, MemoryVAM, Mem-World
system-0 SuperMimic, TAM
agentic-robotics GaP
data-augmentation DAVATaR
dynamic-manipulation TEMPO
generalist π0.7
humanoid SuperMimic
steerability π0.7
submerged-manipulation DAVATaR
system-1 DreamZero
tactile DAVATaR
test-time-training RoboTTT

Themes