FRED-5B-RoboCasa

Checkpoint of FRED (Future REpresentation learning via self-Distillation with temporal context) from Self-Distilling Future Video Representations for World Action Models.

Project page · Code · Paper (arXiv)

FRED trains World Action Models (WAMs) to produce better future video representations from a single video denoising step: the video model is conditioned on a short history of observations, and its intermediate representation of the future is trained by self-distillation from an EMA teacher that sees the noisy ground-truth future observations.


Video backbone	Wan2.2-TI2V-5B
Action decoder	16-layer DiT
Training data	RoboCasa Kitchen (kimtaey/robocasa_mg_gr00t_300)
Training	batch 256, 100K iterations
Files	`checkpoint-100000/` (~33 GB)

Average success rate on RoboCasa Kitchen (from the paper): 75.4%.

Usage

Install the code, then download the checkpoint:

hf download FRED-AGI-LAB/fred-wan22-5b-robocasa-kitchen --local-dir fred-wan22-5b-robocasa-kitchen

The checkpoint directory is fred-wan22-5b-robocasa-kitchen/checkpoint-100000; pass this local path to the scripts (not the repo ID).

Evaluate on every RoboCasa Kitchen task (see docs/robocasa_kitchen.md for the simulator setup and the text cache):

TEXT_EMBED_CACHE=<cache.pt> \
bash run_scripts/eval/eval_fred.sh robocasa_kitchen fred-wan22-5b-robocasa-kitchen/checkpoint-100000 <output dir>

The same directory can warm-start training with BASE_MODEL_PATH=fred-wan22-5b-robocasa-kitchen/checkpoint-100000.

Citation

@article{fred2026,
  title   = {Self-Distilling Future Video Representations for World Action Models},
  author  = {Jang, Huiwon and Jang, Suhyeok and Kang, Hyeonmo and Kim, Seungku and Shin, Jinwoo and Yu, Sihyun},
  journal = {arXiv preprint arXiv:2610.30320},
  year    = {2026}
}