Reproduce “Before the Body Moves: Learning Anticipatory Joint Intent for Language-Conditioned Humanoid Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haozhe Jia, Honglei Jin, Yuan Zhang, Youcheng Fan, Shaofeng Liang, Lei Wang, Shuxu Jin, Kuimou Yu and 4 more
The DAJI framework enables streaming, language-conditioned whole-body control for humanoid robots by learning an anticipatory, dynamics-aligned joint-intent latent as an intermediate representation. The system achieved a 94.42% rollout success rate and superior language-motion alignment on HumanML3D-style benchmarks, demonstrating smoother and more stable transitions in long-horizon streaming scenarios on BABEL.
Key insights
- Explicitly encoding multi-horizon future reference information within a joint-intent latent is critical for achieving anticipatory behaviors and long-horizon stability in dynamic humanoid control.
- Training a low-level controller through in-loop distillation, where the student policy drives its own rollouts and receives guidance from a future-aware teacher, effectively addresses exposure bias and shapes a dynamics-aligned latent space.
- A unified hierarchical framework that decouples semantic intent generation from real-time physical execution, mediated by an anticipatory and executable latent, enables coherent and stable streaming language-conditioned humanoid control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.