Task and Motion Planning (TAMP)

TAMP systems maintain a dual-layer architecture where a symbolic planner reasons about abstract task states while a motion planner validates geometric feasibility.

• Symbolic layer: Uses STRIPS-like operators or PDDL to sequence high-level actions (pick, place, push)
• Geometric layer: Solves inverse kinematics and collision-free trajectories in continuous C-space
• Bidirectional communication: Symbolic plans are grounded by motion feasibility checks; geometric failures trigger symbolic re-planning

This separation allows reasoning about long-horizon goals without getting lost in low-level joint configurations.

Semantic attachments are the critical interface that connects symbolic predicates to continuous geometric solvers.

• A symbolic predicate like Graspable(?obj, ?gripper) invokes an inverse kinematics solver to verify reachability
• Attachments act as oracles that validate or reject symbolic states based on geometric reality
• Common attachments include collision checkers, IK solvers, and grasp stability evaluators
• This mechanism prevents the symbolic planner from exploring geometrically infeasible branches of the search tree

Without semantic attachments, symbolic plans would frequently propose actions that are physically impossible to execute.

TAMP systems explicitly reason about grasp modalities and object affordances rather than treating manipulation as a single action type.

• Power grasps vs. precision grasps: Different contact geometries require different approach vectors
• Regrasping: Planning intermediate placements to reorient objects when direct grasps are occluded
• Tool use: Reasoning about grasped objects as functional extensions for achieving goals
• Non-prehensile actions: Pushing, sliding, and toppling as valid manipulation primitives when grasping is infeasible

This richness enables solving tasks like assembling furniture or clearing cluttered shelves where simple pick-and-place fails.

TAMP employs hierarchical task networks (HTNs) or similar decomposition strategies to break complex goals into manageable subproblems.

• A goal like AssembleGearbox decomposes into InsertShaft, AlignBearing, SeatRetainingRing
• Each sub-task inherits constraints from its parent while introducing local geometric requirements
• Backtracking across hierarchy levels: If a sub-task proves geometrically impossible, the planner can revise higher-level task sequences
• This mirrors how human experts plan: rough sequence first, then detailed motion verification

Hierarchical decomposition is essential for scaling TAMP to real industrial assembly tasks with dozens of steps.

Once a symbolic plan with geometric waypoints is found, TAMP systems apply trajectory optimization to smooth and time-parameterize the motion.

• CHOMP and STOMP: Gradient-based trajectory optimizers that minimize collision costs and smoothness penalties
• Trajectory parameterization: Converting geometric paths into time-stamped joint trajectories respecting velocity and acceleration limits
• Cost functions encode preferences: minimize jerk, maximize clearance from obstacles, minimize execution time
• The optimizer treats the initial sampling-based path as a seed and refines it locally

This two-stage approach—sampling for feasibility, optimization for quality—produces motions suitable for real robot execution.

A defining characteristic of TAMP is unified backtracking across both symbolic and geometric decision levels.

• When a motion planner fails to find a collision-free trajectory for a grasp, the system backtracks to try a different grasp pose
• If all grasps fail, it backtracks further to consider a different object or action ordering
• Conflict-directed search: Failure explanations from the geometric layer inform which symbolic choices to revise
• This contrasts with pipelined approaches where task planning and motion planning are sequential and cannot recover from downstream failures

Unified backtracking is what makes TAMP robust to the inherent uncertainty of real-world manipulation.

TAMP solves the classic warehouse kitting problem by reasoning about object selection order and collision-free grasp trajectories simultaneously.

• Symbolic layer selects which item to pick next based on order requirements and accessibility
• Motion layer computes grasp poses and arm trajectories that avoid bin walls and neighboring objects
• Handles occluded objects by planning push actions to reorient items before grasping
• Reduces cycle time by 30-50% compared to scripted pick-and-place sequences

TAMP generates feasible assembly sequences for multi-part products by reasoning about geometric constraints and insertion order.

• Symbolic planner determines which parts can be assembled next based on precedence constraints
• Motion planner verifies each insertion is physically achievable without collisions
• Automatically discovers regrasping operations when a single grasp cannot complete all insertions
• Applied in automotive powertrain assembly and consumer electronics manufacturing

Combines base navigation with arm motion to enable autonomous mobile manipulators (AMMs) that pick and place items across warehouse zones.

• Task planner sequences navigation goals and manipulation actions
• Motion planner coordinates base and arm trajectories simultaneously for efficient whole-body motion
• Handles dynamic environments where obstacles and inventory positions change
• Deployed in e-commerce fulfillment centers for goods-to-person and person-to-goods workflows

TAMP plans the routing of deformable linear objects like cables and wire harnesses through complex assemblies, a critical task in automotive and aerospace manufacturing.

• Symbolic layer reasons about routing order and which clips or channels to use
• Motion layer computes physically realistic cable deformations using Cosserat rod models
• Plans insertion sequences that avoid tangling and respect minimum bend radius constraints
• Reduces manual routing time by up to 70% in harness assembly cells

TAMP enables bimanual manipulation where two robotic arms collaborate on tasks requiring coordinated motion, such as large-part handling or assembly.

• Task planner decomposes goals into left-arm and right-arm subtasks with temporal dependencies
• Motion planner computes collision-free trajectories for both arms in a shared workspace
• Handles force-controlled interactions like holding a part with one arm while the other fastens
• Used in aerospace panel assembly and heavy equipment manufacturing

TAMP automates the selection and sequencing of tool changes and fixture placements for multi-step machining and fabrication operations.

• Symbolic planner selects appropriate tools for each operation and determines fixture locations
• Motion planner verifies tool access paths and fixture clamping clearances
• Plans intermediate re-fixturing when a single setup cannot provide access to all features
• Reduces setup time and eliminates manual fixture design iterations in CNC machining cells