Counterfactual world models with hindsight planning hit up to 69% success across sim tasks
randall_balestr · x · 2026-11-20
A follow-up post on the counterfactual world models paper by Balestriero et al. shares the experimental side: the team validated counterfactual reasoning across many environments and explored hindsight planning, yielding large success-rate gains.
- Experiments span five rendered sim tasks — HandReach, Ball-in-Cup, FetchReach, BlueROV2, and liquid transport — using JEPA world models over frozen DINOv2-S and MAE-B encoders.
- Paper Figure 11 shows the counterfactual model with outcome readout crushing both the trained predictor and an oracle planner lacking outcome info: 69% on FetchReach, 60% on Ball-in-Cup, 62% on liquid transport, versus 10–25% baselines.
- Conditioning action selection on the observed outcome markedly helps for the same event; code and demos are released.
Related event: LeCun and Colleagues Release Counterfactual World Models Paper(12 posts)→
More from Embodied
- Tesla FSD 15's real-time view isn't 3D rendering — it's the AI model's own synthetic perception — mark_k · 2026-11-22
- China Installed 59% of New Industrial Robots in 2026 Report, ~354,000 Units — ingliguori · 2026-11-22
- Apple Watch bricks after one shower while a $15 Keep band survives many — oran_ge · 2026-11-22
- Tesla FSD V15 visualizations go viral with detailed 3D night-time rendering — DimaZeniuk · 2026-11-22
- Tesla FSD visualization: from dots on a screen to photorealistic 3D night views in six years — XFreeze · 2026-11-22
- YQ Bionics' Y-Hand M2 packs 38 degrees of freedom as AI meets robotic manipulation — CurieuxExplorer · 2026-11-22