HINT-WAM
fine-tuning a world action model for robotics
1 min read
HINT-WAM performs post-training (subsequent fine-tuning) on a world action model called Fast-WAM. During training, the model uses future demonstration videos as hindsight guidance, acting as its own teacher for the action expert via an energy-based objective. At deployment, only the student is used, preserving the base Fast-WAM model's architecture and inference costs. This solution is built for researchers and developers working in robotics and imitation learning.
Key features
- Higher robot control accuracy in the LIBERO and RoboTwin simulators.
- Self-distillation using future demonstration videos (hindsight).
- Full compatibility with the base Fast-WAM architecture and inference settings.
- Ready-to-use checkpoints for integration into existing pipelines.
What you can do
- Use the
libero90/hintwam_libero90.ptcheckpoint for manipulation tasks on the LIBERO-90 benchmark. - Run closed-loop evaluations using standard Fast-WAM emulators.
- Explore the energy-based scoring method for fine-tuning other action models.
Example
from huggingface_hub import hf_hub_download
repo = "tuandattttt/HINT-WAM"
config = hf_hub_download(repo, "config.yaml")
ckpt = hf_hub_download(repo, "libero90/hintwam_libero90.pt")
Read more
DetailsBenchmarks (LIBERO-90): 85.18% success rate (vs. 15.56% for base Fast-WAM and 78.88% for SFT).
Training: AdamW, cosine schedule, batch size 192, 8 GPUs with 24 GB VRAM, bfloat16, DeepSpeed ZeRO-2.
License: MIT.










Comments