750,000+ hours recorded in October!
Models are getting smarter and hardware keeps getting better, but what sits at the core of Physical AI hasn't changed. It still learns from real-world experience.
Context Matters.
What does "put it back" actually mean?
In a kitchen, it's stacking plates, in a garage, it's finding the exact slot for each wrench. Real-world data teaches AI that the same instruction looks different everywhere.
Every skill a robot picks up starts with a person doing it first, whether that's cooking a meal, cleaning a room or repairing a vehicle.
Real data from real people is how robots learn the real world.
CoRL just confirmed what we've seen on the ground. Models aren't what's holding robots back anymore; it's how much of the real world they've actually seen.
CoRL 2026 just wrapped in Austin, and the message from the robot learning community is clear: the bottleneck is no longer the policy architecture. It's data and compute economics.
Context: 2,094 submissions (~3x last year), 687 accepted (32.8%).
The papers I think matter most π
π SuperMimic (UC Berkeley) β Best Paper
Instead of brute-forcing humanoid motion tracking with massive parallel sim RL, it learns one differentiable world model of the dynamics shared across an entire motion library, then trains the policy through short imagined rollouts.
β ~3 hours on a single H100, vs. ~21K GPU-hours (128 GPUs Γ 7 days) for NVIDIA's SONIC
β Fails on 25 of 129,514 motions (baseline: 179)
β Transfers to a real Unitree G1
GRAIL (NVIDIA, UCLA) β Best Paper finalist
Generates 20K+ humanoid loco-manipulation sequences from 3D assets + video priors, without ever filming a person. G1 policies hit 84% pick-up and 90% stair-climbing success.
EgoScale (NVIDIA, UC Berkeley, UMD)
A VLA pretrained on 20,854 hours of action-labeled egocentric human video. Validation loss drops log-linearly with hours (RΒ² β 0.998), a clean scaling law for human-to-robot transfer. +54% avg success on a 22-DoF hand.
DreamZero (NVIDIA)
A 14B world action model running at 7 Hz. Reports >2x generalization vs. leading VLAs, and adapts to a new robot body with ~30 min of play data.
Ο0.7 (Physical Intelligence)
Conditions training on *how* a task was done: speed, a 1β5 quality score, mistakes, control mode. Failed rollouts and interventions become training signal instead of getting deleted.
RoboTTT (NVIDIA, Stanford, et al.)
Test-time training stretches visuomotor context to 8,000 timesteps. The only model tested that finishes a 5-minute, 10-stage assembly.
T-Rex (UC Berkeley, et al.)
Puts ~100 hours of tactile data directly in the action loop: 30%+ higher avg success than the best baseline across 12 contact-rich tasks.
ATHENA
Makes data-influence scoring for a billion-parameter VLA ~313x cheaper, and matches full-data fine-tuning with only half the demos in sim.
My takeaways:
1. Human video is now a real scaling axis, but action labels matter more than raw hours.
2. World models are becoming either the policy itself (DreamZero) or the training engine (SuperMimic).
3. Training compute is collapsing. On-robot inference latency is the new bottleneck.
4. Curation beats collection. Knowing which data to keep (ATHENA, Ο0.7) is worth as much as getting more.
5. Touch is the next data layer.
Caveat: most numbers are author-reported from 2026 preprints and not yet independently reproduced.
Physical AI is shifting from "scale the model" to "engineer the data flywheel."
#CoRL2026 #PhysicalAI #Robotics #RobotLearning
This is where AI really starts.
Behind every model that learns to see, listen and move is someone capturing those experiences firsthand, and here's a look at our community in Vietnam making it happen. π»π³
Our team visited Robotworld 2026 in Korea and Physical AI is moving faster than most people realize.
We saw robots with lifelike faces and capture rigs that see the world through human eyes. The hardware is ready; the next leap will come from the real-world data that teaches robots how humans live, move and interact.
$620,000+ earned by our AI contributors in just 7 days.
AI has been trained on the internet for years without paying the people behind the data.
Thatβs over. Real contributors. Real environments. Real earnings.
Context Matters.
On a factory floor, hand annotation maps every grip, twist and placement as wires are sorted and parts assembled.
The more real work a model sees, the closer it gets to how people move.
@paigeinsf Robotics data is nowhere near solved, and the gap shows the moment a model leaves the lab. Variety, real environments and careful collection are still where most of the hard work is.
Physical AI starts with getting the data right.
Real-world environments. Verified human contributors. Explicit consent.
These are the building blocks of AI that understands how humans interact with the world.
Context Matters.
One kitchen, one task, three views. From the head, you see the whole counter. From the wrists, you see the small details, like how the packet tilts and where the fingers grip.
The more angles a model learns from, the closer it gets to how people really move.