Sequential Mobile Manipulation Planning

• SMMP is a framework that interleaves mobile navigation and manipulation actions, dynamically adapting earlier steps to enable complex sequential tasks.
• It employs unified representations like the Virtual Kinematic Chain and Augmented Kinematic Representation to combine robot, object, and scene information for integrated planning.
• Various architectures balance symbolic sequencing with reactive control, achieving enhanced efficiency and task success in cluttered, dynamic environments.

Searching arXiv for recent and foundational papers on Sequential Mobile Manipulation Planning and closely related mobile-manipulation planning frameworks. Searching arXiv for "Sequential Mobile Manipulation Planning" and the cited papers. Sequential Mobile Manipulation Planning (SMMP) concerns the generation and execution of ordered mobile-manipulation behaviors in which navigation, reaching, grasping, transporting, placing, and articulated-object interaction are interleaved over multiple steps, and in which earlier actions alter the feasibility of later ones. In the literature, SMMP spans several layers of abstraction: symbolic task planning over action sequences, geometric and kinematic planning over coupled base-arm-object states, and execution-layer control that preserves task priorities while reacting to obstacles, humans, and semi-static scene changes (Jiao et al., 2021, Li et al., 2024, Jiao et al., 26 Aug 2025).

1. Problem structure and formal scope

A common formalization models the environment by a state set $\mathcal{S}$ and transitions $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$, with a transition $t=\langle s,s' \rangle$ taking the environment from $s$ to $s'$; the planning problem is to find a sequence of transitions from an initial state $s_0\in S$ to a goal state $s_g\in S_g$, where $S_g\subseteq S$ (Jiao et al., 2021). Within that abstraction, the target task class includes pick-and-place and articulated-object interaction such as opening doors and drawers, i.e. sequential tasks that require navigation, reaching, grasping, transporting, placing, and constrained interaction over multiple steps (Jiao et al., 2021).

The literature does not use a single operational meaning of SMMP. In one line of work, the problem is long-horizon pick-place-move planning for a manipulator on a mobile base in environments with many movable rigid objects, where navigation feasibility depends on prior rearrangement decisions and on whether the robot is holding an object (Garrett et al., 2016). In another, SMMP is posed as continuous execution of a given time-parameterized end-effector trajectory, as in mobile printing or spraying, where feasibility must hold for every time $t$ rather than only at sparse waypoints (Nguyen et al., 2023). A third line treats SMMP as a sequence of trajectory-tracking tasks supplied by a higher-level planner, with the main challenge being reactive whole-body execution under strict task priorities (Du et al., 10 Mar 2026).

This scope variation underlies a recurrent misunderstanding: SMMP is not uniformly a symbolic task-ordering problem. Some papers address full symbolic-to-motion pipelines, while others explicitly address the execution/control layer for a precomputed task sequence (Du et al., 10 Mar 2026, D'Orazio et al., 21 Nov 2025). A second common misunderstanding is that SMMP reduces to alternating “move base, then manipulate.” Several papers argue that this decomposition is often the source of inefficiency or outright failure, especially in clutter, narrow passages, and articulated-object interaction (Jiao et al., 2021, Kang et al., 2018).

2. Unified representations of robot, object, and scene

A major development in SMMP is the shift from decoupled symbolic or kinematic models toward unified representations that consolidate the mobile base, manipulator, and manipulated object into one planning entity. The “Virtual Kinematic Chain” (VKC) perspective constructs a serial virtual chain $\mathcal{C}^V$ by combining the mobile base, the arm, and the manipulated object; the object chain is inverted where needed and attached to the end-effector via a virtual joint, while planar base motion is represented by “two perpendicular prismatic joints and a revolute joint” (Jiao et al., 2021). This removes explicit symbolic commitments to intermediate base poses and replaces conventional navigation/manipulation alternation with abstract mobile-manipulation actions.

In the VKC PDDL formulation, the predicate vocabulary is compressed to (vkcState ?r ?q), (objConf ?o ?s), (free ?v), and (carry ?o ?v), and the domain is organized around three abstract actions: goto-vkc, pick-vkc, and place-vkc (Jiao et al., 2021). The important representational claim is that these are “sufficient to handle various mobile manipulation tasks, from pick-and-place in different setups to foot-arm coordinated and constrained motions (e.g, door/drawer opening)” (Jiao et al., 2021). Symbolically, this removes extra predicates such as HasTool(s), AbleToPick(s), and Reachable(o, s) that arise when the base and manipulator are modeled as separate subsystems.

A later generalization integrates not only robot and manipulated object kinematics but also scene kinematics into an “Augmented Kinematic Representation” (AKR), whose configuration space is the “Augmented Configuration Space” (A-Space) (Jiao et al., 26 Aug 2025). Its motion-planning state is

$\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$0

so the mobile base, robot arm, and articulated scene DOFs are planned jointly (Jiao et al., 26 Aug 2025). This is structurally broader than VKC: doors, drawers, microwaves, and closets become kinematic entities rather than task-specific constraints (Jiao et al., 26 Aug 2025).

Other work reaches similar unification from different directions. “Harmonious sampling” retains low-dimensional base-dominant planning in simple regions but switches to full joint-space sampling in difficult regions identified by a reachability map and narrow-passage analysis (Kang et al., 2018). “SLITHERS” formulates unified state planning for base and manipulator in one Lie-theoretic optimization space, using pose-log constraints on $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$1 rather than separate base placement and arm IK stages (Smith et al., 2024). Across these formulations, the shared claim is that coupled base-arm-object reasoning is most critical exactly where sequential decompositions are weakest.

3. Planning architectures and search regimes

SMMP systems differ sharply in where they place the main computational burden. FFRob represents one influential direction: a tightly integrated task-and-motion planner for long-horizon pick-place-move problems that introduces Extended Action Specification (EAS), allowing arbitrary predicates such as Reachable to appear directly in action conditions (Garrett et al., 2016). This lets heuristic search account for changing geometric and kinematic feasibility under different object placements and grasps, rather than treating symbolic search and motion feasibility as independent layers (Garrett et al., 2016).

By contrast, VKC-based and A-Space methods reduce symbolic complexity by simplifying the domain itself. The dynamic VKC framework organizes planning into task planning, VKC-based motion planning, and reactive MPC; the symbolic planner reasons in PDDL over compact pick and place operators, the motion planner instantiates each action into a whole-body trajectory, and a state monitor triggers replanning after task-relevant state changes (Li et al., 2024). The A-Space framework adopts a tri-level structure: symbolic task planning over AKR evolution, an intermediate plan-refinement stage that selects long-horizon-feasible action goals, and optimization-based motion planning in A-Space (Jiao et al., 26 Aug 2025).

A complementary symbolic line addresses partially known environments through graph representations. EPoG formulates planning as graph transformation on a belief scene graph $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$2, with action sequences derived from graph edit operations between belief and goal graphs and ordered by temporal dependencies and movement costs (Yang et al., 4 Feb 2026). This integrates exploration and sequential manipulation by continuously updating the belief graph with observations and LLM-predicted object placements, then replanning when observations invalidate the current hypothesis (Yang et al., 4 Feb 2026). In a different setting, UniPlan extends tabletop symbolic domains into long-horizon mobile-manipulation domains by adding node-locality predicates such as (robot_at_node ?r ?n) and (object_at_node ?o ?n), door traversal, hand-indexed occupancy predicates, and travel costs on a compressed topological map (Ye et al., 9 Feb 2026).

These differences also clarify a boundary in SMMP research. Some methods choose action order and symbolic subgoals; some assume the sequence and optimize only execution. HTMPC is explicit that it is “mainly an execution-layer planner/controller for SMMP” rather than a solver for task ordering (Du et al., 10 Mar 2026). SM$\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$3ITH retains that position while extending execution to human-centered environments through interactive human prediction and bilevel MPC (D'Orazio et al., 21 Nov 2025). The field therefore spans symbolic sequencing, hybrid symbolic-geometric planning, and reactive sequence execution rather than one canonical pipeline.

4. Motion generation, control, and safety

At the motion level, unified SMMP planners typically solve constrained optimization or sampling problems over whole-body state. In A-Space, the trajectory optimizer minimizes weighted first- and second-order finite differences

$\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$4

subject to chain constraints $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$5, task-goal constraints, joint/velocity/acceleration bounds, and obstacle/self-collision penalties (Jiao et al., 26 Aug 2025). VKC uses a related optimization view in which the terminal goal is specified in a lower-dimensional task space rather than as a fully fixed robot state, leaving coordinated base-arm-object details to the optimizer (Jiao et al., 2021).

Execution-layer methods replace long-horizon geometric search by receding-horizon whole-body control. HTMPC formulates strict task hierarchy as nonlinear lexicographic optimization over a horizon, solving a cascade of single-task MPC problems rather than a weighted-sum controller (Du et al., 10 Mar 2026). On a 9-DoF mobile manipulator, it reports improved hierarchical trajectory tracking performance by $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$6 on average over a task-prioritized inverse-kinematics baseline under task changes, singularity, and reference variations, and a hierarchical task control architecture that executes a sequence of delivery tasks $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$7 faster than a typical single-task architecture (Du et al., 10 Mar 2026). The reported implementation uses $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$8, horizon $\mathcal{T}\subseteq\mathcal{S}\times\mathcal{S}$9, $t=\langle s,s' \rangle$0, average compute time $t=\langle s,s' \rangle$1, and operation at $t=\langle s,s' \rangle$2 (Du et al., 10 Mar 2026).

Safety and adaptivity become central once SMMP is executed in dynamic or changing environments. The dynamic VKC framework separates changes with long-term symbolic consequences from short-term local disturbances: the former trigger online task replanning, while the latter are handled by reactive MPC tracking a nominal VKC trajectory (Li et al., 2024). SM$t=\langle s,s' \rangle$3ITH extends hierarchical MPC with interactive human prediction through bilevel optimization and adds discrete-time control barrier function constraints $t=\langle s,s' \rangle$4 for safety in human-centered environments (D'Orazio et al., 21 Nov 2025). Perceptive HTMPC adds object-level Bayesian scene-change inference and map-derived control barrier functions to maintain a temporally accurate 3D environment representation during long-duration sequential tasks; it runs the whole-body MPC at $t=\langle s,s' \rangle$5 Hz, sends commands at $t=\langle s,s' \rangle$6 Hz, and reports an average closed-loop delay around $t=\langle s,s' \rangle$7 s (Du et al., 10 Mar 2026).

A persistent misconception is that reactive obstacle avoidance alone solves dynamic SMMP. The safety-oriented literature does show strong execution benefits from multi-sensor workspace monitoring, obstacle prediction, elastic-band adaptation, and online slowdown/stop behavior (Frese et al., 2020). However, the dynamic SMMP literature distinguishes local geometric disturbances from state changes that invalidate the remaining symbolic plan, such as a cabinet being closed by a human after the robot had opened it (Li et al., 2024). In that sense, reactive control is necessary but not sufficient for sequential mobile manipulation.

5. Applications and empirical evidence

The empirical record is heterogeneous because the papers evaluate different subproblems, but several quantitative patterns recur. First, unified symbolic abstractions reduce search depth and branching. In VKC-based object rearrangement with $t=\langle s,s' \rangle$8 objects on $t=\langle s,s' \rangle$9 tables, planning time and generated search nodes grow much more slowly than with a conventional domain, and for the 16-object problem the reported empirical averages are $s$0 and $s$1 in the search-space scaling argument over 50 trials (Jiao et al., 2021). In the later AKR/A-Space formulation, a rearrangement study reports an average $s$2 reduction in search depth for 16 objects across 50 trials (Jiao et al., 26 Aug 2025).

Second, unified base-arm-object planning improves feasibility in cluttered manipulation. In simulation studies for A-Space planning, the paper reports an 84.6\% higher task success rate compared to baseline methods, and real-robot validation across seven types of rigid and articulated objects in 17 contexts demonstrates long-horizon tasks of up to 14 sequential steps (Jiao et al., 26 Aug 2025). The dynamic VKC framework further demonstrates successful sequential task execution under human interference by combining online PDDL replanning with whole-body VKC planning and reactive MPC (Li et al., 2024).

Third, selective or reduced-dimensional coupling can be computationally advantageous when full joint planning is too expensive. Harmonious sampling reports improvements “up to 5.6 times faster” in time to find an initial solution and “up to 17% lower” in final solution cost over a decoupled approach, especially in difficult scenes with narrow space (Kang et al., 2018). For continuous-process mobile manipulation, MoboConTP plans the base trajectory first in configuration spacetime and then recovers arm IK, with formal completeness and optimality for the discretized base-planning problem and planning-time comparisons such as 89.9 s versus 120.3 s at 8 layers and 155.1 s versus 515.0 s at 14 layers against a Dijkstra baseline (Nguyen et al., 2023).

Fourth, exploration-integrated and language-conditioned systems show that SMMP in partially known indoor environments benefits from compact symbolic abstractions rather than generic frontier exploration. EPoG reports a success rate of $s$3 across 46 realistic household scenes and 5 long-horizon daily object transportation tasks, with an average travel-distance reduction of $s$4 (Yang et al., 4 Feb 2026). UniPlan, in turn, programmatically extends local manipulation domains into building-scale mobile-manipulation domains over a visual-topological map and is evaluated on 50 human-raised tasks with door traversal, hand occupancy, and travel cost in the symbolic model (Ye et al., 9 Feb 2026).

6. Assumptions, boundaries, and open problems

Despite their diversity, current SMMP formulations share several strong assumptions. VKC-based symbolic planners assume pre-sampled feasible object configurations and grasping poses, known static environment geometry, and contact represented mainly through attach/detach semantics rather than rich contact-state reasoning (Jiao et al., 2021). Continuous-task planners such as MoboConTP assume a given time-parameterized end-effector trajectory, known base velocity bounds, and reachability-based admissible sets, but do not include manipulator joint velocity/acceleration constraints or guarantees of continuous joint-space arm motion (Nguyen et al., 2023). Execution-layer methods such as HTMPC and SM$s$5ITH assume that the task sequence is already provided and optimize only a short active window at a time (Du et al., 10 Mar 2026, D'Orazio et al., 21 Nov 2025).

Another recurring boundary concerns uncertainty and environment change. EPoG handles unknown object locations through symbolic belief-graph revision and LLM-guided semantic hypotheses rather than probabilistic belief-space planning (Yang et al., 4 Feb 2026). UniPlan assumes a pre-built visual-topological map and sufficiently accurate image-to-symbol grounding (Ye et al., 9 Feb 2026). Perceptive HTMPC and SM$s$6ITH push farther toward semi-static change and human-centered interaction, but they still rely on specific perception and motion-prediction models rather than formal robust stochastic guarantees (Du et al., 10 Mar 2026, D'Orazio et al., 21 Nov 2025).

A third limitation is contact richness. Much of the SMMP literature covers prehensile manipulation and articulated-object interaction with known kinematics, but not general non-prehensile manipulation, forceful contact, or dynamic interaction. H-MaP shows that sequential manipulation with dynamic contact-mode switches, tool use, and auxiliary object handling can be planned by hybrid waypoint generation, contact sampling, and optimization-based motion planning, but it is centered on a fixed-base Franka Panda rather than full mobile manipulation (Cicek et al., 2024). Likewise, the multi-mobile-manipulator flipping framework addresses sequential segmentation, coordinated platform motion, and regrasping under closed-chain constraints, but assumes a given object trajectory and predefined grasp sets rather than solving a general long-horizon mobile-manipulation problem end to end (Liu et al., 2023).

The trajectory of the field suggests three persistent open directions. One is tighter integration of symbolic long-horizon reasoning with continuous feasibility, beyond post hoc motion feasibility checks or shallow lookahead refinement. A second is robust execution under perception error, scene drift, and human interaction without sacrificing the efficiency gains of compact symbolic domains. A third is broader interaction modeling: richer contact transitions, non-prehensile actions, uncertainty-aware planning, and real-robot autonomy without external infrastructure remain only partially addressed in the surveyed work (Li et al., 2024, Du et al., 10 Mar 2026, Jiao et al., 26 Aug 2025). These limitations suggest that SMMP remains best understood not as a single algorithmic template, but as a family of methods for sequencing, instantiating, and safely executing coupled navigation-manipulation behavior over long horizons.