pickplace_full_v1

running

pi0.5 full fine-tune batch 32 lr 5e-5 ema 0.999 horizon 50 A100 80GB

Progress

Step

Flow-matching loss

Log scale — the value spans 101×. One loss is all pi0.5 reports; the vision-language backbone is trained by gradient flow from this same MSE.

latest

Gradient norm

Log scale. Bounded and settling is healthy; a sustained climb is not.

Parameter norm

Linear, narrow band. Rising means the weights are actually moving.

Table view — every logged point

Step	Epochs	Loss	Grad norm	Param norm	Checkpoint

Reading this

The loss log lags the run. Telemetry stops at step while the run is at — Python block-buffers stdout when it is piped, so tee receives the lines late. Add PYTHONUNBUFFERED=1 to see it live.

The absolute value reads low by construction. 11 of the 32 action dimensions are constants — 7 left-hand joints commanded to exactly 0, plus 4 padding dims — so roughly a third of the mean squared error is over channels with nothing to predict. Compare the curve to itself, never to another dataset.

The curve has stopped discriminating. Over the last 2,000 logged steps the mean loss moved — smaller than the step-to-step noise. Every checkpoint from here on is indistinguishable by loss, which is not the same as equivalent on the robot.

So pick by rollout, not by this page. Two checkpoints with equal loss can differ by tens of points of success rate, and nothing plotted here measures grasp success. Choosing among the surviving checkpoints needs real-robot trials from a fixed object-placement grid.