MotionDisco: Humanoid Loco-Manipulation Discovery

• MotionDisco is a hybrid framework that discovers long-horizon, contact-rich humanoid loco-manipulation maneuvers from scratch without human demonstrations.
• It integrates an LLM-guided evolutionary search with sequential kinodynamic trajectory optimization to efficiently generate and validate complex contact plans.
• The system trains reinforcement-learning tracking policies on discovered trajectories, achieving zero-shot hardware transfer across diverse benchmark tasks.

MotionDisco is a framework for discovering long-horizon, contact-rich humanoid loco-manipulation motions from scratch, without teleoperation, human demonstrations, or motion retargeting. It treats extreme humanoid behavior as a coupled discrete-continuous problem: an LLM-guided evolutionary search proposes interaction sequences over contacts, objects, and support transitions, while a sequential kinodynamic trajectory optimizer checks and realizes those sequences as dynamically feasible whole-body trajectories. The discovered trajectories are then used to train reinforcement-learning tracking policies that transfer zero-shot to a real humanoid robot. The work presents this as the first system to discover and deploy long-horizon humanoid loco-manipulation skills entirely through automated evolutionary search (Taouil et al., 4 Jun 2026).

1. Problem scope and formalization

MotionDisco studies what the paper terms extreme humanoid loco-manipulation: tasks in which a humanoid must combine locomotion, manipulation, whole-body support interactions, and long-horizon reasoning. The evaluated scenarios include climbing onto a table while moving a box, carrying an object across clutter, stepping onto and down from elevated surfaces while holding objects, using objects or furniture as intermediate supports, and reaching into constrained spaces under a table (Taouil et al., 4 Jun 2026).

The central difficulty is not only the dimensionality of humanoid motion, but the combinatorial explosion in possible contact interactions. Over a long horizon, the robot must choose which hand or foot contacts which surface, whether an object remains supported or has been lifted, when to climb, when to relocate obstacles, and in what order these events occur. The feasibility of one decision can depend on later ones, so the hard part is often to discover an appropriate contact plan before solving for a trajectory.

The framework formalizes a contact mode as a complete assignment over contact interfaces:

$$c = \{(a,b)\ \forall a \in I : b \in I \cup \{\emptyset\}\},$$

with the symmetry condition that if $a$ is in contact with $b$, then $(b,a)$ must also appear in the mode. For a contact mode $c$, the feasible manifold is defined as

$$M(c) := \left\{ (x,\dot{x},u)\ \middle|\ \begin{array}{l} \text{dynamics}(x,\dot{x},u,c) = 0, \ \text{contact}(x,u,c) \le 0, \ \text{collision}(x) \le 0, \ \text{limits}(x,u) \le 0 \end{array} \right\}.$$

This decomposition makes explicit that contact planning and motion realization are related but distinct subproblems.

The task input consists of a scene representation $S$ and a goal $G$. The goal may include discrete target contacts and continuous target poses. A plausible implication is that MotionDisco is designed less as a generic controller and more as a planner-discoverer for structured, multi-stage, environment-dependent skills.

2. Contact-plan programs and LLM-guided evolutionary search

MotionDisco searches over contact-mode sequences, or equivalently interaction sequences, but it does not represent them as unstructured symbol strings. Each candidate is an executable program that emits a contact plan. The program uses helper functions such as get_initial_mode(), walk(plan, n_steps), and append_mode(). This representation compresses ordinary locomotion into a walk() primitive and reserves explicit symbolic editing for contact-relevant events such as grasps, lifts, placements, step-ups, and support releases (Taouil et al., 4 Jun 2026).

The prompt exposes a structured contact vocabulary. A mode contains channels such as left_ee, right_ee, left_sole, right_sole, and one channel per movable object. End-effector and sole entries take tuples of the form $(\text{env\_id}, \text{contact\_type})$ or free; object entries take $$(\text{self\_id}, \text{env\_id}, \text{contact\_type})$$ or free. The only supported contact type is unilateral. The prompt also imposes task-specific physical rules, including that at least one sole must remain in contact at all times, that an object must be held with both end-effectors before its support breaks, and that both hands must remain in contact while carrying until the object is placed again.

The LLM used for mutation and plan generation is Claude Opus 4.7. Its role is not low-level control; it operates at the level of contact-sequence reasoning. The prompt includes scene geometry, surface IDs, helper-function documentation, example transition patterns, and geometric heuristics. Among the built-in heuristics are that walk() performs straight-line locomotion, that movable obstacles may need relocation before a walk that would collide, that the robot can reach roughly $a$0 m from the base, and that the step-up height limit is about $a$1 m.

Search proceeds as an LLM-guided evolutionary tree search. Each node stores a program $a$2, its score $a$3, optional text feedback $a$4, and an offspring count $a$5. The tree is partitioned into islands. Within island $a$6, node $a$7 is weighted by

$a$8

where $a$9, $b$0 is the island median score, and $b$1 is a median-based scale. This biases search toward promising nodes while discouraging over-expansion of the same lineage.

Mutation types are sampled with probabilities

$b$2

These correspond to localized edits, full rewrites, and crossover. Novelty filtering rejects near-duplicate candidates using embedding similarity and, when needed, an LLM novelty judge. This suggests that MotionDisco treats diversity of contact strategies as a first-class search objective rather than an accidental byproduct.

3. Sequential kinodynamic trajectory optimization and pruning

Once a candidate contact program is executed, MotionDisco evaluates the resulting contact-mode sequence with a two-stage planner. The first stage is a sequential kinematic feasibility check that ignores full dynamics but tests geometric realizability, collision avoidance, and transition consistency. The second stage is a full contact-explicit trajectory optimization that solves for robot-object state trajectories, controls, and stage durations only for candidates that survive the kinematic filter (Taouil et al., 4 Jun 2026).

For a fixed contact sequence $b$3, the continuous optimization problem is

$b$4

subject to

$b$5

$b$6

and

$b$7

The optimizer uses direct multiple shooting. The sequential kinematic stage is solved with acados, and the full trajectory optimization stage uses HIPPO.

Candidates that reach full trajectory optimization are scored by

$b$8

where $b$9 collects equality and inequality constraint residuals at knot point $(b,a)$0, and $(b,a)$1 is a jerk regularizer. Lower trajectory-optimization cost is better.

A decisive efficiency mechanism is longest-prefix feasibility analysis. If a candidate fails the kinematic stage, MotionDisco computes

$(b,a)$2

It then reports the first failing mode or transition as textual feedback to the LLM. This converts planner failure into structured search guidance. In practical terms, the system does not merely reject a bad plan; it localizes where the discrete interaction strategy breaks.

4. Policy training and hardware transfer

The trajectories discovered by MotionDisco are not executed directly by the optimizer. Instead, they serve as reference motions for reinforcement-learning tracking policies trained in simulation and deployed on hardware (Taouil et al., 4 Jun 2026).

The policy-training stage uses DeepMimic-style rewards and domain randomization. The paper does not provide an explicit reward equation in the main text; the tracking objective is described at the level of reference-motion following with robustness sufficient for zero-shot real-robot transfer. The important architectural point is that the system separates motion discovery from control execution: the search-and-optimization stack discovers feasible long-horizon trajectories, and RL absorbs those trajectories into executable feedback policies.

This division of labor has two consequences. First, MotionDisco does not require teleoperated demonstrations or retargeted human motion even at the policy-learning stage; the references come from automated search. Second, the method can exploit robot-specific affordances rather than remain constrained by human-like demonstration priors.

The paper reports real-robot deployment on tasks including Under-Table Pick & Place, Move Through Clutter, Parkour Pick & Place 2, Long-Distance Pick & Place, and Climb Table w/ Box. It emphasizes multiple consecutive successful executions across all attempted tasks.

5. Empirical evaluation

MotionDisco is evaluated on eight benchmark tasks: Banana, Box Stacking, Climb Table w/ Box, Long-Distance Pick & Place, Move Through Clutter, Parkour Pick & Place 1, Parkour Pick & Place 2, and Under-Table Pick & Place (Taouil et al., 4 Jun 2026). These tasks are designed to stress multi-stage contact reasoning rather than isolated locomotion or isolated manipulation.

The quantitative comparison uses three regimes: SC (single LLM call), MD w/o TF (MotionDisco without text feedback), and MD (full MotionDisco with text feedback). A contact plan is counted as valid if it passes the sequential kinematic feasibility check and trajectory optimization converges. The paper reports that single-call generation is often insufficient, that both MotionDisco variants solve all scenarios, and that full MotionDisco generally yields a higher percentage of valid plans and lower trajectory-optimization cost than the no-feedback version.

The reported time to first valid solution for the full method is only a few minutes across all tasks:

Task	Time to first valid solution
Banana	7.49 min
Box Stacking	4.58 min
Climb Table w/ Box	2.15 min
Long-Distance Pick & Place	2.50 min
Move Through Clutter	3.11 min
Parkour Pick & Place 1	6.05 min
Parkour Pick & Place 2	4.40 min
Under-Table Pick & Place	1.34 min

A notable qualitative result is that the search does not collapse to a single solution family. For Parkour Pick & Place 2, the paper shows multiple valid contact plans with different hand and foot trajectories in a single run. This suggests that the island structure, novelty filtering, and mutation operators preserve distinct high-level strategies rather than merely refining one seed.

The ablation evidence supports the framework design. Single Call often fails completely. MotionDisco without text feedback remains capable, but the planner-feedback loop improves both efficiency and solution quality. The combined picture is that iterative search matters, and structured textual feedback from the planner materially improves the discrete search over contact sequences.

6. Limitations and open directions

The current framework has several explicit restrictions (Taouil et al., 4 Jun 2026). Its contact model supports only unilateral sticking contacts, rectangular patches, and rigid box-like manipulated objects. It assumes a known scene representation rather than integrating perception, and it relies heavily on prompt-embedded priors and helper abstractions such as straight-line walk().

These choices are not incidental. They are part of what makes the search space tractable. At the same time, they bound the current scope. Likely difficult cases include articulated furniture, sliding or pivoting contacts, non-box objects, scenes where straight-line walking is inadequate, and tasks requiring online scene inference rather than pre-specified surface IDs. The paper also notes that more complex scenes with more objects and surfaces will preserve the underlying combinatorial character of the problem.

Future directions identified in the paper include extending the contact model to sliding and non-unilateral contacts, supporting articulated and free-form objects, integrating perception/vision to build scene representations from observations, and exploiting emerging simulation-ready scene-construction pipelines for on-the-fly motion discovery. A plausible implication is that MotionDisco’s current contribution is primarily algorithmic: it establishes a workable search-and-realization template for long-horizon humanoid loco-manipulation under structured contact assumptions.

7. Distinction from similarly named “DisCo/Disco” systems

MotionDisco is unrelated to several other systems with similar names. DisCo is a diffusion-based framework for controllable human dance image/video synthesis that disentangles subject appearance, background, and pose (Wang et al., 2023). Disco4D is a Gaussian-splatting framework for disentangled 4D human generation and animation from a single image (Pang et al., 2024). Disco-LoRA studies multi-concept video customization through disentangled composition of content, style, and motion in text-to-video models (Xu et al., 25 Jun 2026). Another DisCo addresses visual encapsulation in video MLLMs through semantic distinctness and temporal coherence of visual tokens (Zhao et al., 14 Jul 2025).

MotionDisco also differs sharply from music-conditioned dance-generation systems. DiffDance is a cascaded diffusion model for music-to-3D-dance generation (Qi et al., 2023), and TempoMoE structures a diffusion transformer around tempo-aware mixture-of-experts for rhythm-aligned dance synthesis (Lyu et al., 21 Dec 2025). Those systems generate dance motion from music. MotionDisco instead discovers and deploys whole-body humanoid loco-manipulation trajectories through discrete search over contact interactions coupled to kinodynamic planning (Taouil et al., 4 Jun 2026).

The shared lexical motif therefore obscures a substantive methodological difference. In current usage, MotionDisco denotes a hybrid symbolic-continuous motion discovery system for real humanoid robots, not a diffusion-based content generator, not a video customization model, and not a music-conditioned dance framework.