After 600 experiments, here's what I discovered about JEPA world models
Web page · https://haddyalchaer.com/jepa
First seen by Oriel on .
AI summary
After 600 LeWM experiments on Push-T, the report finds that planning changes—not better predictions—deliver major gains for long-horizon and out-of-distribution goals without retraining.
- LeWM’s learned predictor achieved 78% success versus 77% with simulator physics on 250 long-horizon tasks following trained-path distributions.
- Splitting 75-step tasks into 25-step subgoals increased success from 22% to 80%, because each search stayed within LeWM’s effective planning horizon.
- A memory graph of roughly 460,000 embedded demonstration states generated subgoals, reaching 85% after adding costs to jumps between demonstrations.
- Shifting, rotating, and mirroring demonstration segments restored performance from 32% to 85% after deleting goal-nearby memories, while using only 1% of demonstrations.
- Scoring only the block instead of the pusher quadrupled exact-physics success on harder goals, while flexible goal matching doubled LeWM’s success from 11% to 23%.
#jepa #world models #robot planning #push-t
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 292, latest observation ; +286 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 49, latest observation ; +48 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50487185 | https://haddyalchaer.com/jepa Read discussion |