2x speed

This program was written by a coding agent

No supplied planner, skills, or simulator source code

Coding Agents for Generalized Task and Motion Planning ProblemsAnonymous ICRA 2027 submission

TAMP planner

Generalized TAMP

Coding agent

engineer / LLM

predicates

operators

samplers

skills

abstractions

environment

coding agent

planner

planner

planner

generalized policy

program

TAMP problems with a variable number of objects

TAMP planner: given abstractions, one search per instance

Generalized TAMP: given abstractions, one policy for all instances

Drop the abstractions: only a coding agent and the environment?

task description

# Shelf3DEnv (variable object count)
A 3D task where the robot must pick up objects from the ground and place them onto a space-constrained shelf in a cupboard with three layers. The robot has a holonomic mobile base with powered casters and a Kinova Gen3 arm.
This environment contains a VARIABLE number of objects.

simulator client

4x

stalls on the floor · its recalled arm model says the fingertip is 2.4 cm above it

8x

where its prior arm model put the block · shown 5x

arm model error, 41 probes: 38.9 mm → 1.8 mm

6x

drops the cube at the shelf and never goes back for it

3x

fix: checks each grasp with its arm model, retries cubes left on the floor

reset(seed) · step(action) · render()
object-centric state in, action out

sandbox

Python · NumPy · SciPy
no environment source · no network
$20 · 60 s per test instance

coding agent

approach.pyfrozen · no LLM

15x speed

100 held-out instances, 1 to 8 cubes
no LLM at test time
0 solved

Given: a task description, a simulator, a bare sandbox and $20

It probes: the arm model it wrote from memory is wrong

It holds a block and fits its own arm model

It writes a policy: reset() and get_action()

It tests on the seeds it chooses: four of five fail

It debugs and commits, 13 times in this run

Frozen program, 100 held-out instances, no LLM

task description

# StickButton2DEnv (variable object count)
A 2D environment where the goal is to touch all buttons. The robot has a movable circular base and a retractable arm with a rectangular vacuum end effector.
This environment contains a VARIABLE number of objects. Your program must handle ANY number of them.

simulator client

reset(seed) · step(action) · render()
object-centric state in, action out

sandbox

Python · NumPy · SciPy
no environment source · no network
$20 · 60 s per test instance

coding agent

approach.pyfrozen · no LLM

real time

100 held-out instances, 1 to 5 buttons
no LLM at test time
0 solved

Given: a task description, a simulator, a bare sandbox and $20

It probes the black box: read states, push until the wall stops it

It writes a policy: reset() and get_action()

It tests where it wants: 300-seed sweeps, worst cases first

It debugs and commits, 14 times in this run

Frozen program, 100 held-out instances, no LLM

28 environments × 5 synthesis methods × 5 runs each × 100 held-out instances

70,000 evaluation episodes

14 of 28 environments at 95% success or above

Mean success per family

AgenticGenPlanCoding agent, bare sandbox, no abstractions. One program per environment.

PlannerHand-written predicates, operators, samplers, skills. New plan per instance.

LLMGenPlanLLM writes the program from the environment source. Fixed feedback, no tools.

AgenticGenPlan · main setting · 0% Claude Code

AgenticGenPlan + source · 57% Claude Code

6x speed

Policy computation per action · 44 matched seeds, 13 environments

With the environment source: recovered, 7x slower per action

AgenticGenPlan · main settingClaude Code

3x speed

Goalevery cube in the bin

human demonstration · 4x

expectedunderhand throw

Winds up to an overhand throw

AgenticGenPlan · main settingClaude Code

8x speed

Goalget the balls to the right of the wall, at least half of them

human demonstration · 4x

expectedsmall scoops

Regrasps the tool, then one huge scoop

AgenticGenPlan + sourceClaude Code

5x speed

Goalmove the pile of crumbs into the green tray

human demonstration · 4x

expecteduse the scoop tool

Grabs and dumps the tray, ignoring the tool

AgenticGenPlan + sourceClaude Code

6x speed

Goalopen the drawer and sweep the crumbs in

human demonstration · 4x

expecteduse the sweep tool

Sweeps with the arm instead of the tool

AgenticGenPlan · main settingClaude Code

4x speed

Goalevery block inside the shelf

Makes room by pushing stored blocks deeper

AgenticGenPlan · main settingClaude Code

4x speed

Goalretrieve the target block to the target region

25 objects: clears a path to the target

AgenticGenPlan + sourceClaude Code

6x speed

48x speed

Goaleach cube in the bin of its colour

Recovers from failed grasps

AgenticGenPlan + sourceClaude Code

real time

Goalpush the movable button onto the target button

Swings the hook like a windmill, pushing with its long side

AgenticGenPlan · main settingCodex · early version of the program

real time

Goalany blue block on the green plate

Early in synthesis: the grasp fails

AgenticGenPlan · main settingCodex · final program

1.5x speed

Goalany blue block on the green plate

Later: it goes for the spare blocks on the far table

0.0 s captions space play · arrows cue · shift+arrows scene · t theme · c captions · ?t=SS ?theme=x ?record=1