Goaliphy

Your robot failed. Find the data that caused it.

Goaliphy traces failed rollouts back to the training episodes behind them, then retrains to check that the fix works.

Send us your failures Start with 5–10 failed rollouts
Failed rollout Wrong object

Asked to put the bowl on the plate, the policy picked up the mug.

Training episodes, ranked by influence on success hurts ← → helps
0214
−0.31

Reason: the instruction names the bowl, while the gripper closes on the mug.

0188
−0.24
0097
−0.17
0142
+0.11
0301
+0.08
Fig. 1. Illustrative example. A failed rollout traced to the training episodes behind it.
Why filters aren't enough

Correct data can still make a policy worse.

In a published experiment, Pebblous added 93 physically correct synthetic episodes to a 299-demonstration GR00T training set, without reweighting. Closed-loop success fell. Open-loop prediction error stayed at about 0.001 in both runs, so the offline number showed nothing.

A filter that checks each episode alone can't see harm that comes from the mixture. Rollout outcomes can.

Closed-loop success rate
299 demonstrations
73%
Same, plus 93 correct synthetic episodes
43%
Fig. 2. Data from Pebblous, July 2026. Not a Goaliphy result. Read their report
How it works

From a failure to a fix you can verify.

  1. 1

    Start with the failure

    Each rollout is logged with what went wrong and the checkpoint that produced it.

  2. 2

    Find the episodes

    Attribution ranks the training episodes that pushed that failure.

  3. 3

    Get a written reason

    Every flag says why, as in Fig. 1, so your team reviews a list and nothing disappears silently.

  4. 4

    Retrain to confirm

    We retrain without the suspects and test whether the policy really changed.

One dataset, different answers

The best data for speed may not be the best data for reliability.

One overall score hides trade-offs. Goaliphy scores each episode against each metric, so you can curate for the outcome you care about.

Episode HRT RMST F(30 s) Hard failures
0088 +0.8 +0.5 −0.6 −0.7
0301 −0.6 −0.5 +0.7 +0.8
0142 +0.6 +0.5 +0.6 +0.7
0214 −0.8 −0.7 −0.7 −0.9
0057 +0.1 +0.2 0.0 +0.1
0176 +0.5 +0.6 −0.4 −0.3
0269 −0.3 −0.2 +0.5 +0.6
0033 −0.1 +0.1 −0.2 0.0
Fig. 3. Illustrative example. Episode 0088 helps speed and hurts reliability. Episode 0301 does the reverse.
Helps Hurts
Pilot

Send 5–10 failures. We trace them, then prove the fix by retraining.

Built for robotics teams without an in-house data and evaluation group.