VQV Signal

ROBOTS & HARDWARE WATCH TECHNICAL

CriticHack study finds reward model trade-off in robot drawer task

According to a post on X by @dustinturnerai, a Nov 21 rundown covers several recent robotics papers. In the revised CriticHack study, a visual reward model raised a real robot's correct drawer opening from 10.8% to 17.5%, but wrong-drawer failures rose from 8.3% to 24.2%, across 120 trials per condition. The TacInno work reports learning which contact changes vision cannot explain, averaging 70.5% success on four physical tasks versus 61.5% for the best tactile baseline. The post's title also lists SAFI on safe interruption and RynnValue-8B in LeRobot, but the snippet cuts off before giving details on those.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-22T05:40:00+00:00 Detected 2026-11-22T12:03:46+00:00
View original source

According to a post on X by @dustinturnerai, a Nov 21 rundown covers several recent robotics papers. In the revised CriticHack study, a visual reward model raised a real robot's correct drawer opening from 10.8% to 17.5%, but wrong-drawer failures rose from 8.3% to 24.2%, across 120 trials per condition. The TacInno work reports learning which contact changes vision cannot explain, averaging 70.5% success on four physical tasks versus 61.5% for the best tactile baseline. The post's title also lists SAFI on safe interruption and RynnValue-8B in LeRobot, but the snippet cuts off before giving details on those.

AI-assisted summary based on the listed source.

@dustinturnerai's Nov 21 rundown: in the revised CriticHack study, a visual reward model raised a real robot's correct drawer opening from 10.8% to 17.5%, but wrong-drawer failures rose from 8.3% to 24.2% (120 trials per condition). TacInno learns which contact changes vision cannot explain, averaging 70.5%...

For robot builders, the CriticHack figures suggest a reward model can improve the target behavior while also increasing a different kind of error, so evaluations should track wrong actions as well as successes. The TacInno result points to touch sensing as a possible complement to vision in manipulation. These are summaries from a single social post, so readers should check the original papers for methods and conditions before relying on the numbers.

Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.

Signal Strength 88% Technical label WATCH Public Interest 24 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 18 Curiosity Score 48 Shareability Score 21

VQV surfaced this signal because it is recent, relevant to AI Search, connected to Scobleizer News Pulse (X).