Same robot.
Different instruction.
Different winner.
General-purpose models are learning to act. Explore where PEAR's evidence supports that bet, and where specialist policies still lead.
What changed?
Success rate (%). PEAR paper Tables 2–3; π0.5 values cited from prior work. Different training and compute prevent an architecture-only conclusion.
How often does the goal change?
A hypothetical mix of standard tasks and changed-goal tasks. Move the slider to weight the two published results.
Crossover ≈ 28.4% in this arithmetic model. This is not a measured deployment threshold.
18 successes. Keep the other 12 in view.
Every published trial of the final checkpoint. Select a tile to see the task, outcome and round count.
Two trials per scene/object: these observations are clustered. The independent-trial interval is illustrative, not a guarantee for unseen tasks. Final model practiced corresponding task types in RoboDojo, but not this real robot.
Failure ledger
6 ran out of rounds · 3 finished unaligned · 2 toppled the object · 1 ball rolled away.
221 tokens a second.
How long before the next move?
Animation repeats the calculated decode interval. Actual robot actions are not simulated.
Formula: reasoning tokens ÷ tokens per second. This combines a BF16 reasoning-length measurement with NVFP4 throughput for sensitivity only. NVFP4 also changes reasoning length and accuracy. Prefill, vision, network and movement time are excluded.
Robot use could become
a general model capability.
My bet is on general intelligence gaining physical experience. VLAs and WAMs already use foundation models and remain valuable approaches. This experiment examines the evidence without declaring the problem solved.