Panda-TAMP — Closed-Loop Task & Motion Planning

PDDL symbolic planning, bridged to real geometry, executed on a simulated Franka Panda — with a planner that knows when it's wrong and replans.

I built this project to answer a question that most robot-planning demos dodge: what happens when the symbolic plan meets geometry that disagrees with it? Classical task planners happily emit pick(cup); place(cup, tray) with no idea whether the cup is reachable, graspable, or whether the tray is 2.5 metres away. Most demos hide that gap. This project makes the gap the whole point: the planner and the geometry engine argue with each other in a closed loop until they agree on a plan that is both logically correct and physically executable — or until the system honestly reports that no such plan exists.

flowchart LR
    subgraph Perception
        CAM[RGB-D camera] --> YOLO[YOLOv8 detector] --> BP[depth back-projection] --> WS[/world_state/]
    end
    subgraph Planning["Closed TAMP loop"]
        WS --> PG[PDDL problem generator]
        PG --> FD[Fast Downward<br/>symbolic plan]
        FD --> GR[Geometric refiner<br/>IK · grasp sampling · swept-path collision]
        GR -- "infeasible: falsify bridge predicate" --> PG
        GR -- feasible --> PLAN[/task_plan/]
    end
    subgraph Execution
        PLAN --> EXE[MoveIt 2 / OMPL executor]
        EXE --> GZ[Gazebo Harmonic<br/>physics grasping]
        GZ -- "execution failure -> falsify + replan" --> PG
    end
    subgraph Language["Language layer (optional)"]
        NL["natural-language command"] --> GG[goal grounder<br/>LLM or grammar] --> PG
        MCP[MCP server] --> PG
    end

The core idea: bridge predicates

PDDL plans purely symbolically (pick / place / stack / unstack). Geometry — kinematics, grasps, collisions — lives in a separate engine (pykin FK + a damped-least-squares IK I wrote myself, plus trimesh/FCL collision checking). The two layers meet through four geometric bridge predicates:

reachable(o) · graspable(o) · placeable(o, s) · stackable(o, o')

The symbolic planner treats them as ordinary facts, but it never gets to decide them — the geometric layer does. The loop:

1. Seed optimistically. Every bridge predicate starts true.

2. Plan symbolically. Fast Downward produces a candidate action sequence.

3. Refine geometrically. For each action: sample grasps, solve IK, sweep the gripper along the actual joint-space path, and check the arm links against the scene — not just the endpoint poses.

4. On failure, falsify and loop. The offending predicate is set false in the PDDL initial state and the planner runs again — now knowing that fact.

5. Execution failures feed the same loop. A grasp that slips or a placement that misses at runtime falsifies the same predicates and triggers a replan from the observed world state, not the assumed one.

The part I like most: the symbolic planner needs no geometric knowledge and the geometric engine needs no search. Each does the thing it is good at, and the falsification protocol is the entire interface between them.

Honesty is a design constraint

A recurring failure mode of planning demos is quietly reporting success. I built the opposite bias into every layer:

• The replanning benchmark asserts the property that matters (the block can only ever land on the reachable tray) instead of a scripted replan count — and it includes a counterfactual test proving the solver reports failure when no feasible surface remains, rather than emitting a plan onto one it can't reach.

• The Gazebo executor verifies arrival by TF (a plan is not an execution), gates every grasp on measured finger closure (a weld with open fingers is not a grasp), and runs a drop detector against simulator ground truth.

• The reliability harness (scripts/reliability_pnp.sh) scores a run as a pass only if the plan completed and Gazebo's ground-truth poses show every object physically on the target surface. Log-parsing alone doesn't count.

What's inside

Quickstart

The core is deliberately ROS-free — the full plan⇄refine⇄replan loop runs in plain Python in seconds:

python3 -m virtualenv -p python3.12 .venv && . .venv/bin/activate
pip install -e ".[dev]"

tamp-offline --scene stack3     # symbolic + geometric plan for a 3-block stack
tamp-offline --scene replan     # watch the loop veto an unreachable tray
tamp-command "put the cups on the tray" --scene pnp   # language -> goal -> plan
python -m pytest tests/unit -q  # 70+ tests, no ROS required

Example output (replan scene — the goal offers two trays, one physically unreachable):

=== TAMP [replan] ===
success=True  replans=0  reason=ok
  1. PICK(block, ...)                -> EEF [0.45  0.    0.435]
  2. PLACE(block, s=tray_near, ...)  -> EEF [0.55  0.2   0.453]
[replan] counterfactual (execution later falsifies tray_near too):
[replan]   success=False ... falsified=[('placeable','block','tray_near'), ('placeable','block','tray_far')]
[replan]   -> honest failure: no feasible tray remains.

Full stack (ROS 2 Jazzy + MoveIt 2 + Gazebo Harmonic)

cd ros2_ws && colcon build --symlink-install && cd ..
bash scripts/run_ros_full.sh stack3      # plan + MoveIt execution
bash scripts/run_ros_full.sh replan      # closed-loop recovery over ROS topics
bash run.sh pnp                          # Gazebo physics pick-and-place (GUI)
bash scripts/reliability_pnp.sh 6        # ground-truth-verified reliability batch

The ROS layer runs two Python interpreters (the planner node needs the pinned geometry stack; MoveItPy needs the system ROS python) — scripts/ros_env.sh and the launch files handle the split. Setup details and the non-obvious MoveItPy / Gazebo integration notes are in docs/RUNNING.md and docs/ros_integration.md.

Language-grounded goals & the MCP server

The part I find most fun: you can drive the planner conversationally, and it still can't be talked into doing something infeasible.

$ tamp-command --scene pnp "put all the cups on the tray"
  goal [rule]: {'type': 'sort_by_class', 'assignment': {'cup_1': 'tray', 'cup_2': 'tray'}}
  ✓ plan (4 steps, 0 replan):
     1. PICK cup_1
     2. PLACE cup_1
     3. PICK cup_2
     4. PLACE cup_2

$ tamp-command --scene pnp "do not put cup_1 on the tray"
  ✗ refused: the command is a prohibition, not a goal

The grounder chain is LLM-optional: Anthropic API → OpenAI-compatible (NVIDIA NIM) → local Ollama → a deterministic grammar that always works offline. Every backend's output — LLM or not — passes through the same validating DSL and the same planner. The model proposes; the planner verifies. There is also an MCP server exposing get_world_state / plan_goal / plan_command as tools, so any MCP-capable client or agent can read the robot's world and submit goals under exactly the same verification contract.

Details: docs/nl_grounding.md.

Repository layout

tamp/                 ROS-free core
├── pddl/             domain model + problem generator (bridge-predicate seeding)
├── planning/         Fast Downward wrapper
├── geometry/         robot model (DLS IK), grasp sampling, collision, refiner
├── reasoning/        pose -> symbolic facts (on / stacked-on / clear)
├── execution/        execution-failure -> falsified-predicate mapping
├── nl/               goal DSL, grammar + LLM grounders
└── mcp_server.py     the planning stack as MCP tools
pddl/domains/         canonical tabletop.pddl
benchmarks/           scene fixtures (stack3, replan, pnp) — also test fixtures
goals/                example goal specs (YAML)
ros2_ws/src/          ROS 2 overlay: planner node, MoveIt executor, grasp manager,
                      YOLOv8 + mock perception, launches, worlds, custom msgs
scripts/              runners: offline demo, full ROS bringup, reliability harness
tests/                unit + integration (the core runs anywhere Python runs)
docs/                 architecture, running guide, ROS notes, language layer

Testing

python -m pytest tests/unit tests/integration -q

70+ tests cover the domain encoding, the reasoner, the refiner (including swept-path and per-link collision regressions), the closed loop (recovery and honest failure), the NL grammar's safety refusals, DSL validation, and the MCP tool surface. CI runs the ROS-free suite on every push — deliberately only that: I don't fake Gazebo coverage in a container that can't run it; the physics-dependent layer is exercised by the ground-truth reliability harness on a real machine instead.

Engineering notes & honest limitations

Things I learned the hard way, and the edges I have not sanded off:

• Simulated grasping is its own research problem. The gripper uses Gazebo's DetachableJoint with finger-closure gating and contact-band verification — a documented compromise, not force-closure physics. Under Gazebo Harmonic's bullet-featherstone backend the fixed-joint constraint is violable (the upstream fix lands in a later Gazebo generation), which bounds multi-object pick-and-place reliability; the reliability harness exists precisely to measure that honestly rather than cherry-pick green runs.

• Physics backends disagree. DART silently drops URDF mimic joints; bullet enforces articulated dynamics faithfully enough to expose bad inertials and contact-margin effects that DART masked. The gripper stack (independent finger joints, effort limits, collision geometry, joint limits) is tuned against measurements, and the measurement scripts ship in scripts/.

• The refiner's arm model is deliberately coarse — a swept gripper proxy plus per-link spheres. It catches gross pass-throughs cheaply; MoveIt's full-mesh planning scene remains authoritative at execution time.

• Simulation only. No physical robot yet: the execution layer speaks standard MoveIt 2 / ros2_control interfaces, so a real Panda is the natural next step, but I make no hardware claims until it runs on one.

Roadmap

• Force-aware grasping (effort-interface gripper control) to retire the detachable-joint compromise entirely

• The upstream fixed-constraint enforcement once it ships in binaries (or a local backport) for higher multi-object reliability

• Real-hardware bring-up on a physical Panda

• Richer domains: obstructions, regrasping, multi-step manipulation puzzles

License

MIT — see LICENSE.