Technical Summary: MotionDisco

Problem Statement

Humanoid robots require the ability to execute long-horizon loco-manipulation behaviors in unstructured environments to operate effectively in the real world. However, exploring high-dimensional, contact-rich spaces is combinatorially complex and extremely challenging. Current state-of-the-art approaches typically rely on motion retargeting (transferring human demonstrations to robots) or teleoperation to bypass the exploration problem. These methods face inherent limitations: generating demonstrations for every possible scenario is impractical, and tying motions to human data constrains the robot to human-like behaviors, preventing the exploitation of the robot's full physical capabilities. Furthermore, recovering accurate contact information from kinematic retargeting alone is difficult, often requiring dynamic refinement before reinforcement learning (RL) tracking.

Methodology: MotionDisco Framework

The paper proposes MotionDisco, a framework that discovers contact-rich, long-horizon humanoid loco-manipulation motions from scratch without teleoperation or human demonstrations. The system couples a Large Language Model (LLM)-guided evolutionary search over discrete contact plans with an efficient sequential kinodynamic trajectory optimizer.

1. Contact-Explicit Motion Planning

The motion planning module evaluates candidate contact sequences through a two-stage process:

• Sequential Kinematic Feasibility: A fast geometric filter that rejects contact sequences failing collision avoidance, configuration limits, or transition consistency. It returns text feedback identifying the specific failure point.
• Trajectory Optimization (TO): For kinematically feasible sequences, a continuous optimization problem solves for state trajectories, controls, contact wrenches, and stage durations. This stage enforces dynamics, contact constraints, and hardware limits. The resulting trajectory provides a cost metric (constraint residuals + jerk regularization) and structured feedback.

2. LLM-Guided Evolutionary Tree Search

The search operates over a tree of contact-mode sequences, represented as executable programs:

• Representation: Each node contains a program that generates a sequence of contact modes (assignments of end-effectors and objects to surfaces).
• Mutation: An LLM (specifically Claude Opus 4.7) mutates parent programs based on the scene description, task goal, parent score, and text feedback from the motion planner. The LLM generates new contact plans by editing the code, focusing on discrete contact assignments (grasps, lifts, placements, step-ups) while leaving continuous motion recovery to the optimizer.
• Search Strategy: The framework employs a hierarchical population model (inspired by ShinkaEvolve) to maintain diversity. It uses island-based sampling to distribute search effort, novelty-based rejection to avoid redundancy, and "inspiration sampling" to allow new candidates to build on diverse prior solutions.
• Feedback Loop: The motion planner's feedback (e.g., "IK infeasible at mode 5") is translated into natural language and fed back to the LLM, guiding subsequent mutations to repair local failures rather than discarding entire sequences.

3. Real-World Deployment

Once a successful trajectory is discovered, a reinforcement learning tracking policy is trained using DeepMimic-style rewards and domain randomization. This policy is then deployed zero-shot on a real humanoid robot to execute the discovered motions.

Key Contributions

1. Novel Framework: A system coupling LLM-guided evolutionary search over contact plans with a layered kinodynamic optimizer. The optimizer provides both numerical costs and structured language feedback, enabling time-efficient discovery of complex behaviors.
2. Search Efficacy: Extensive ablation studies demonstrate that the iterative search with structured feedback significantly outperforms non-iterative approaches (single LLM calls). The evolutionary nature of the search yields diverse motions for the same task in a single run, providing an efficient source of data generation.
3. Real-World Validation: The work presents the first autonomous discovery and execution of long-horizon humanoid loco-manipulation skills on real hardware without motion retargeting or teleoperation.

Results

The framework was evaluated on eight tasks ranging from simple pick-and-place to complex whole-body behaviors (e.g., climbing tables while carrying objects, navigating clutter, and parkour-style pick-and-place).

• Success Rate: MotionDisco successfully solved all scenarios, whereas a single LLM call (without iterative search) failed to find valid solutions for several tasks.
• Efficiency: The framework found the first valid solution within a few minutes across all scenarios.
• Cost Reduction: The evolutionary search with text feedback achieved lower trajectory optimization costs and higher percentages of valid contact plans compared to a version without text feedback.
• Diversity: The search produced qualitatively distinct contact plans (e.g., different end-effector assignments) for the same task, demonstrating the system's ability to explore the solution space broadly.
• Hardware Execution: Discovered trajectories were successfully transferred to a real humanoid robot, achieving multiple consecutive successful executions across all attempted tasks.

Significance and Claims

The paper claims that MotionDisco represents a significant step forward by enabling humanoid robots to reason about their environment and synthesize novel long-horizon behaviors entirely through automated evolutionary search. It moves beyond the limitations of imitation learning and teleoperation, demonstrating that robots can unlock physical capabilities beyond human-like behaviors. The work establishes that combining the reasoning capabilities of LLMs with rigorous kinodynamic optimization and structured feedback is a viable path for discovering complex, contact-rich skills in high-dimensional spaces.

Limitations

The authors acknowledge current limitations:

• The contact model supports only unilateral sticking contacts on rectangular patches.
• Manipulated objects are restricted to rigid, box-like geometries.
• The system assumes a known scene representation and does not yet incorporate real-time perception (e.g., vision modules) to build scene representations from sensor data.
• Future work aims to extend the contact model to sliding/non-unilateral contacts, handle articulated/free-form objects, and integrate on-the-fly scene construction.