DreamDojo
A research paper by NVIDIA on a robotic world model built on Cosmos 2.5, focusing on world modeling for robotic manipulation scenarios. The paper received an ICML Spotlight recognition but was subsequently found to contain multiple serious bugs in its code.
Timeline (last 90 days)
DreamDojo is a robotic world model paper built by NVIDIA on top of Cosmos 2.5.
The DreamDojo paper received an ICML Spotlight honor.
Researchers identified serious bugs in DreamDojo's pre-training, post-training, and evaluation code.
The DreamDojo research team collected approximately 44,000 hours of human data for pre-training.
DreamDojo was trained using 256 H100 GPUs.
A researcher discovered a critical bug in DreamDojo's post-training code with the assistance of Claude.
Table 4 in the paper shows that DreamDojo's PSNR metric improvement over Cosmos 2.5 is only approximately 0.5 dB.
The above content regarding bugs in DreamDojo's code originates primarily from a single whistleblower's account on Reddit and has yet to be verified by NVIDIA officially or by independent reproducers.
All Facts (8)
The above content regarding bugs in DreamDojo's code originates primarily from a single whistleblower's account on Reddit and has yet to be verified by NVIDIA officially or by independent reproducers.
50%UnverifiedTable 4 in the paper shows that DreamDojo's PSNR metric improvement over Cosmos 2.5 is only approximately 0.5 dB.
50%UnverifiedDreamDojo is a robotic world model paper built by NVIDIA on top of Cosmos 2.5.
50%UnverifiedResearchers identified serious bugs in DreamDojo's pre-training, post-training, and evaluation code.
50%UnverifiedThe DreamDojo research team collected approximately 44,000 hours of human data for pre-training.
50%UnverifiedDreamDojo was trained using 256 H100 GPUs.
50%UnverifiedThe DreamDojo paper received an ICML Spotlight honor.
50%UnverifiedA researcher discovered a critical bug in DreamDojo's post-training code with the assistance of Claude.
50%