Task Setting
We evaluate the framework on dexterous manipulation tasks in simulation. The main tasks include DexArt Bucket and Laptop, with an additional Spray Bottle task used to test the method on a new object interaction setting.
Goal Representation Ablation
We compare different 3D hand goal representations on the Bucket task. This ablation studies the trade-off between compactness and expressiveness. A compact goal is easier for the high-level policy to predict, while a richer goal can provide more detailed guidance for finger placement and contact formation.
Effect of Goal Conditioning
We compare policies with and without explicit goal conditioning. Goal conditioning provides the low-level controller with a structured 3D target for the hand, which helps guide contact formation and object manipulation.
References
[1] Bao, Chen, Helin Xu, Yuzhe Qin, and Xiaolong Wang. “DexArt: Benchmarking Generalizable Dexterous Manipulation with Articulated Objects.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. https://arxiv.org/abs/2305.05706