Summary
According to a post on X by @Humanoidguide, AIRoA-0 averaged 78% success across 10 tasks after training on more than 16,000 hours of human recordings and no robot data. Separately, Song Han's team says its open PEAR model went from 0 to 18 of 30 real-robot tests after being trained only in simulation, using the authors' own numbers. The headline describes PEAR as an open 27B model that learns arms from simulation. Both approaches aim to reduce reliance on robot data, which the source describes as a bottleneck.
AI-assisted summary based on the listed source.
What happened
Humanoid.guide reports AIRoA-0 averaged 78% success across 10 tasks after training on more than 16,000 hours of human recordings and no robot data. Separately, Song Han’s team says its open PEAR model went from 0 to 18 of 30 real-robot tests trained only in simulation (the authors’ own numbers); both point to...
Why it matters
Robotics teams that depend on costly robot-collected data could gain alternative training routes if these results hold up. The PEAR figures come from its authors, so independent testing is the main thing to watch. Next, look for published details on task setup and whether the results repeat on other robots.
What this means for you
Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.
Signal Intelligence
Signal Strength 91%
Technical label WATCH
Public Interest 29
Category ROBOTS & HARDWARE
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 0
Novelty Interest Score 70
Consequence Score 30
Curiosity Score 84
Shareability Score 21
Why this is here
VQV surfaced this signal because it is recent, relevant to Robotics, connected to Scobleizer News Pulse (X).
