Nvidia's DreamDojo world model wins ICML spotlight despite critical bugs in released code
Amazing-Fox-7295 · reddit · 2026-10-08
A Reddit post alleges serious flaws in DreamDojo, Nvidia's robotics world model built on Cosmos 2.5, which was accepted as an ICML spotlight.
The story
- Well-known authors; 44k hours of human data (not open-sourced) for pre-training; 256 H100 GPUs used
- Paper's Table 4 shows only 0.5 dB PSNR improvement over Cosmos 2.5 — suspiciously marginal given the resources
- The poster failed to reproduce results via post-training on the released GR1 data; a colleague, using Claude, found a bug that invalidates the entire post-training code
- GitHub issues reveal two more reported bugs affecting the whole pre-training stage (e.g. issue #23)
- Net: pre-training, post-training, and evaluation all appear buggy per the released code, which would explain the paper's weak gains
The author asks how neither the team nor reviewers flagged the massive-effort-minimal-gain red flag. Discussion on OpenReview.
More from Embodied
- Humanoid robots now walk in 3-inch Louboutin high heels, a brutal balance test — kscottz · 2026-10-08
- RAI Institute's Roadrunner robot blends wheels and legs with a single unified policy — Olivier__OG · 2026-10-08
- Burkov: a humanoid robot that can't do this with its hands is useless — burkov · 2026-10-08
- KingKong Robotics open-sources software toolkit for its six-legged crab robot Jumper — kevin_zakka · 2026-10-08
- PhysEvo: recursive self-improvement around a frozen model hits 62% on RoboDojo, 84% on real robot — Wenqing Tian · 2026-10-08
- Developer uses Codex for CAD and build docs in physical robot hardware project — OpenAIDevs · 2026-10-08