- The paper introduces a fully autonomous method for contact-rich, long-horizon humanoid control using LLM-guided evolutionary search combined with kinodynamic trajectory optimization.
- It details a dual-module approach that iteratively refines Python-based contact plans through structured language feedback and dynamic feasibility checks.
- The framework achieves state-of-the-art performance on diverse benchmarks and enables zero-shot deployment on real robots, removing the need for human motion priors.
LLM-Guided Motion Discovery for Long-Horizon Whole-Body Humanoid Control
Introduction
The presented paper introduces MotionDisco, a novel framework for fully automated discovery of contact-rich, long-horizon humanoid loco-manipulation behaviors without reliance on human motion retargeting or teleoperation (2606.06139). The central innovation is the combination of a LLM (LLM, Claude Opus 4.7) guided evolutionary program search with a layered kinodynamic trajectory optimizer. This approach addresses a primary challenge in humanoid control: efficient exploration through the colossal combinatorial space of multimodal contacts and environmental interactions over extended task horizons. MotionDisco enables rapid synthesis of whole-body behaviors for previously unseen, complex manipulation and locomotion tasks, deployable on real robots.
Methodology
MotionDisco operates through two tightly coupled modules: an LLM-guided evolutionary search over discrete contact plans, and a hierarchical trajectory optimization pipeline. The framework iteratively proposes, mutates, and evaluates candidate contact-mode sequences represented as executable code, leveraging structured feedback from geometric and dynamic feasibility checks to refine the search.
The contact planning pipeline begins with rapid, sequential kinematic screening akin to [25], discarding contact sequences failing collision, reachability, or transition consistency. Survivors undergo direct multiple-shooting trajectory optimization subject to full rigid-body dynamics, contact complementarity (unilateral sticking contacts), physical limits, and collision avoidance, yielding dynamically feasible motions for downstream tracking policy training.
Contact modes are modeled as assignments over all contact interfaces (hands, feet, object supports), requiring consistency to avoid unmodeled constraint violations. For a fixed contact-mode sequence, the optimal control problem seeks the state, control, and mode durations that minimize a composite objective (constraint residuals + control jerk), with explicit switching times and final-state goals.
LLM-Guided Evolutionary Program Search
MotionDisco represents each contact plan as a Python program generating a mode sequence. A population-based evolutionary tree search (with ShinkaEvolve-inspired hierarchical/island population structure [22]) drives global and local exploration. Parent node selection is based on node scores, offspring counts, and diversity metrics.
Mutations—ranging from localized diffs, full rewrites, to crossovers—are performed with the LLM, conditionally prompted on the parent program, structured feedback from failed attempts, and inspiration samples from the search tree. Novelty-based filtering prevents convergence to redundant behaviors, explicitly encouraging motion diversity.
Critically, when infeasibility is detected (kinematically or dynamically), the optimizer identifies the failing transition, returning precise language-based feedback to the LLM to guide local repair rather than global rejection—a mechanism shown essential in ablation for efficient discovery and cost minimization.
Experimental Evaluation
MotionDisco is evaluated across eight long-horizon loco-manipulation benchmarks varying in complexity: stacking, parkour-style climbing, long-distance transportation, and manipulation in clutter or confined spaces. All scenarios require integration of reasoning over contact sequences and geometric/dynamic feasibility, without human priors.
Numerical Findings
- Iterative search with text-based structured feedback dramatically outperforms single-call LLM or feedback-free iterative approaches. Across all tasks, the framework discovers a higher percentage of valid contact plans and achieves lower trajectory optimization cost (constraint violations + jerk), with first valid solutions found within minutes.
- The evolutionary nature of the search produces diverse motion policies for a single task in the same run, directly addressing the mode collapse typical in program synthesis with LLMs.
- Discovered motions are successfully tracked with RL-based policies and deployed zero-shot on a real humanoid robot, marking the first demonstration of fully autonomous, non-demo-based, long-horizon loco-manipulation in hardware.
Analysis and Discussion
MotionDisco’s primary contribution lies in disentangling skill acquisition from the need for human demonstration and retargeting, enabling robotic behaviors exceeding the anthropomorphic motion regime. By structuring exploration as LLM-guided code evolution, the system scales efficiently in extremely large, discrete contact manifolds, exploiting informative, actionable feedback from optimization failures. Unlike previous approaches limited to tabletop manipulation, simple assembly, or those leveraging only human-analogous priors [6,10,13,19], MotionDisco generates complex, compositional contact sequences involving support, object relocation, and nontrivial concurrent contacts supporting real-world deployment.
Qualitatively, this evolutionary program search framework broadens the horizon for fast, autonomous generation of training datasets for RL—critical for overcoming data scarcity in robot learning. The algorithm’s ability to generate motion diversity in a population setting further fuels generalization and robustness, crucial for control under uncertainty.
Limitations and Future Work
The framework’s current design restricts contact models to unilateral, sticking contacts on polyhedral surfaces and assumes rigid, box-like object geometries. Extension to articulated, non-polyhedral, or sliding contacts would substantially enhance the class of solvable tasks. Furthermore, perception is not end-to-end: a structured scene description is assumed rather than reconstructed from robot sensory input. Inclusion of vision-based scene parsing [29,30,31] could enable on-the-fly motion discovery in unstructured environments, closing the sim-to-real gap for open-world deployment.
Conclusion
MotionDisco constitutes an advance in autonomous contact-rich skill discovery for humanoid robots, demonstrating the practical potential of combining LLM-guided code synthesis with feedback-driven optimization for scalable, long-horizon motion generation. By removing dependencies on human motion priors, the approach enables task-centric, environment-adaptive behaviors. Its success on nontrivial, real-world manipulation and locomotion tasks points toward future integration with richer contact models and perception, foreshadowing the automated evolution of whole-body, generalized capabilities in embodied systems (2606.06139).
