Imagine trying to solve a complex maze. If you start from the entrance, you are immediately confronted with an overwhelming number of branching paths, most of which lead to dead ends. However, if you start from the center—the goal—and trace your way outward, the paths often converge cleanly toward the exit.
This core asymmetry is the foundation of a powerful paradigm in robotics and manufacturing: assembly-by-disassembly. When a robot needs to construct a complex object, find a collision-free path for a tight-tolerance component, or coordinate multi-agent construction, searching forward from an empty workspace is computationally daunting. By flipping the problem—simulating the deconstruction of a completed target object and reversing the resulting trajectory—planners can bypass the combinatorial explosions and geometric bottlenecks that plague forward search.
The Fundamental Asymmetry of State Spaces
Why is taking something apart easier than putting it together? The answer lies in how geometric and physical constraints manifest at different stages of the planning horizon.
In an empty workspace, a robot attempting forward assembly faces an enormous branching factor. Any given part could theoretically be placed in countless orientations and locations. Conversely, when starting from a fully assembled product, the parts are highly constrained by their neighbors. Only a small subset of components can be removed without causing immediate collisions.
This insight dates back to early robotics research. Masakazu Ejiri et al. demonstrated in An Intelligent Robot with Cognition and Decision-Making Ability that calculating a disassembly sequence and inverting it dramatically simplifies the planning process for complex structures. By formalizing this state-space search via AND/OR graphs, Luiz Homem de Mello et al. showed in And/Or Graph Representation of Assembly Plans that backward search inherently filters out invalid combinations early on. Because every valid disassembly step corresponds to a reversible assembly operation, the backward path guarantees feasibility while exploring a fraction of the state space.
Escaping the Funnel Problem
In contact-rich manipulation tasks—such as inserting a peg into a hole or locking a notched joint—robots encounter what is known as the funnel problem or the narrow-passage problem.
Forward Assembly (Hard): [Open Space] ======> O (Tight Socket)
Backward Disassembly (Easy): [Tight Socket] O ======> [Open Space]
As a robot guides a part into a tight socket, the volume of collision-free configurations shrinks drastically. Randomly sampling a successful entry trajectory in forward time is like searching for a needle in a haystack.
By reversing the problem, the physics work in the robot's favor. As Zih-Yun Chiu et al. explain in Bimanual Regrasping for Suture Needles using Reinforcement Learning for Rapid Motion Planning, initiating a motion planner at the goal state and searching backward allows the trajectory to move from a highly constrained state into free space, where the probability of sampling collision-free paths increases sharply.
The authors of Assemble Them All: Physics-Based Planning for Generalizable Assembly by Disassembly leverage this property within a physics simulator to automatically discover complex insertion motions, such as unscrewing bolts or navigating intricate joints. Similarly, Bingjie Tang et al. demonstrate in AutoMate: Specialist and Generalist Assembly Policies over Diverse Geometries that finding an exit path from a tight-tolerance socket is computationally trivial compared to finding an entry path, making reversed trajectories a highly efficient mechanism for generating robust insertion policies.
Automated Supervision for Learning Agents
Modern reinforcement learning (RL) and imitation learning algorithms excel at contact-rich tasks but are severely bottlenecked by the sparse reward problem. If an agent only receives a reward upon successful assembly, it may never discover the correct sequence of motions through random exploration.
Disassembly acts as an automated data engine that generates high-quality training signals without requiring human demonstrations. In BulletArm: An Open-Source Robotic Manipulation Benchmark and Learning Framework, Dian Wang et al. show that recording the deconstruction of a target structure and reversing the trajectory provides a clear, step-by-step roadmap that mitigates sparse rewards. This strategy is highly scalable; Yijie Guo et al. demonstrate in SRSA: Skill Retrieval and Adaptation for Robotic Assembly Tasks that automated disassembly trajectories provide an effective supervisory signal for representation learning, allowing robots to master the underlying physics of contact-rich tasks.
This deconstructive approach can even be used by agents to generate their own visual instructions. In Learning to Build by Building Your Own Instructions, Aaron Walsman et al. introduce InstructioNet, an agent that uncovers the internal structures and hidden connections of unfamiliar objects by taking them apart first. The reversed sequence of disassembly images serves as an explicit visual blueprint for reconstruction.
Furthermore, this principle applies to complex geometric reasoning. Xinghao Zhu et al. show in Multi-level Reasoning for Robotic Assembly: From Sequence Inference to Contact Selection that simulating disassembly forces along the inertia axes of parts yields physically grounded, collision-free training data. For irregular objects, such as shattered fragments, Yan Shen et al. demonstrate in BiAssemble: Learning Collaborative Affordance for Bimanual Geometric Assembly that training models on reversed simulated disassembly trajectories allows dual-arm robots to discover collision-free fitting sequences based purely on the geometric relationships of the fractured parts.
Grounding Long-Horizon Spatial and Physical Constraints
When moving from individual part insertion to large-scale construction—such as 3D printing, multi-agent building, or clutter manipulation—forward planning frequently suffers from reachability dead ends. A robot building a structure from the ground up might inadvertently place a component that blocks its own arm from reaching the next insertion point, or it might build a sub-structure that collapses under gravity before it can be secured.
Planning in reverse naturally resolves these long-horizon constraints by treating future spatial requirements as immediate boundaries for earlier actions.
Forward Construction: [Build Foundation] ===> [Add Pieces] ===> [Robot Trapped / Collision]
Reverse Planning: [Full Structure] ===> [Remove Piece] ===> [Workspace Opens Up]
As Hejia Zhang et al. demonstrate in A MIP-Based Approach for Multi-Robot Geometric Task-and-Motion Planning and Multi-robot geometric task-and-motion planning for collaborative manipulation tasks, resolving continuous motion parameters backward from the goal configuration ensures that early robot movements do not occupy spaces necessary for completing subsequent phases of a task.
In the domain of additive manufacturing and spatial extrusion, Caelan Reed Garrett et al. show in Scalable and Probabilistically Complete Planning for Robotic Spatial Extrusion that planning a sequence by removing elements from a finished structure progressively clears the robot's workspace. This ensures maximum clearance during the most crowded, highly constrained final stages of construction.
Physical feasibility is similarly preserved through reverse planning:
- Stability: In ASAP: Automated Sequence Planning for Complex Robotic Assembly with Physical Feasibility, Yunsheng Tian et al. frame sequence planning as a disassembly problem to efficiently isolate stable, collision-free removal trajectories.
- Structural Support: For multi-directional 3D printing, Chenming Wu et al. demonstrate in General Support-Effective Decomposition for Multi-Directional 3-D Printing that reversing a simulated segment-removal sequence guarantees that every newly deposited layer has an existing foundation to support it, eliminating the need for sacrificial support structures.
- Traversability: The authors of Multi-agent Collective Construction using 3D Decomposition simulate deconstruction to verify that components can be removed without violating gravitational constraints or blocking agent pathways.
- Collision-Free Growth: Even in abstract geometric settings, Nada Almalki et al. prove in Growing Fast without Colliding: Polylogarithmic Time Step Construction of Geometric Shapes that simulating the collapse of a target shape back into a single seed node provides a mathematically sound, collision-free path for rapid parallel assembly.
By systematically leveraging the geometric openness that emerges when moving away from a goal state, assembly-by-disassembly transforms intractable, long-horizon forward searches into highly efficient, physically grounded planning sequences.
Key Foundational Work
An Intelligent Robot with Cognition and Decision-Making Ability (1971) — contributed 2 concept elements to this direction.
And/Or Graph Representation of Assembly Plans (1986) — contributed 1 concept elements to this direction.