University of North Carolina at Chapel Hill

“DexCompose achieves 77.4% average composite success rate across 16 task combinations”

“The paper demonstrates that two independently trained manipulation policies can be composed at inference time without modifying either base policy. The core formulation keeps both pretrained policies frozen and learns only lightweight "residual" correction modules on top.”

“The LLM-based selector achieves 73.0% mean success vs. 66.5% for the heuristic baseline (Table 3), with the critical example being that the LLM chooses a lower-retention grasp that frees the index finger for downstream manipulation”

“The Task-A stabilizer trains in ~20 minutes on a single RTX 4090 with 1024 parallel environments; the Task-B residual takes ~4 hours (Appendix B.6).”

“CoorDex: Coordinating Body and Hand Priors for Continuous Dexterous Humanoid Loco-Manipulation”

“The zero success rate shows that monolithic latent prediction does not produce reliable grasping under the same reward budget”