JustRUA improves robot-agent scores by changing the harness, not training a new policy
JustRUA improves robot-agent scores by changing the harness, not training a new policy
A November 11 preprint tests the software surrounding GPT-6 Astra rather than introducing another learned robot policy. The authors report an evaluation score rising from 29.7 to 50.3 in RoboDojo simulation. The retained changes preserve episode history, shorten motion chunks, supply robot readings and predicted depth, and let the model calculate coordinates. Research record.
The number needs its denominator and its caveats. The study uses 111 episodes across 42 tasks, not the complete official evaluation. Each configuration is run once; task weighting differs from the official protocol. The authors explicitly say these are not leaderboard scores. Full study.
Guth News analysis
The useful finding is methodological: before crediting a stronger model for a robot’s improvement, separate the model from its observations, action interface and execution loop. The wrapper can alter what the system has a chance to notice and correct.
That is also why this is not evidence that general-purpose robot deployment is solved. A simulator removes many of the consequences that make hardware testing expensive. The next credible comparison would hold the interface and evaluation constant, repeat runs, and test whether gains survive real sensors and physical disturbances.
For builders, the immediate lesson is to measure changes individually rather than stack clever components and attribute the result to intelligence.
Evidence and chronology
Event and primary publication: 2026-11-11 21:08:31 UTC. First covered in the Nov14 01:53 UTC run (Nov13 PST). No independent replication or independent corroborating reporting verified.
Evidence limits: one model, one simulation benchmark, single runs per configuration; no real-world deployment or independent replication verified.
Sources and citations
Revision history
Published .