clutteredstorage medium oracle by weiqianwang123 · Pull Request #88 · tomsilver/robocode

weiqianwang123 · 2026-03-30T22:46:12Z

initial ,working fine for test

…d stuck in pushing the block in shelf

Copilot

Pull request overview

Adds a new “ClutteredStorage2D-b3” medium oracle (approach + behaviors + helpers) plus accompanying tests and artifact handling.

Changes:

Introduces ClutteredStorage2DOracleApproach and StoreRemainingBlocks behavior for the kinder/ClutteredStorage2D-b3-v0 task.
Adds geometry/observation parsing helpers and action/path planning utilities for the oracle.
Adds unit tests (including optional video recording) and updates ignored test artifact directories.

Reviewed changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
tests/oracles/clutteredstorage2d_medium/test_store_remaining_blocks.py	Regression test for `StoreRemainingBlocks` behavior (optionally records video).
tests/oracles/clutteredstorage2d_medium/test_approach.py	Oracle approach tests for known seeds + a solve-rate reporting test.
tests/oracles/clutteredstorage2d_medium/test_approach_debug.py	Debug-style test that writes per-step logs and rollout video artifacts.
tests/oracles/clutteredstorage2d_medium/init.py	Package marker for the new test module.
src/robocode/oracles/clutteredstorage2d_medium/obs_helpers.py	Observation layout parsing + geometry helpers for shelf/blocks/robot.
src/robocode/oracles/clutteredstorage2d_medium/behaviors.py	Main oracle behavior implementation and planning logic.
src/robocode/oracles/clutteredstorage2d_medium/approach.py	Approach wrapper that exposes the behavior through `BaseApproach`.
src/robocode/oracles/clutteredstorage2d_medium/act_helpers.py	Path planning and waypoint→action conversion helpers.
src/robocode/oracles/clutteredstorage2d_medium/init.py	Package marker for the new oracle module.
.gitignore	Updates ignored directories for generated test artifacts.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-01T19:50:53Z

tests/oracles/clutteredstorage2d_medium/test_approach.py

+def test_oracle_solve_rate_report():
+    """Report solve rate across a fixed seed set without gating on 100% yet."""
+    render_mode = "rgb_array" if MAKE_VIDEOS else None
+    shared_env = KinderGeom2DEnv(ENV_ID) if not MAKE_VIDEOS else None
+


test_oracle_solve_rate_report runs 100 seeds with up to 800 steps each but only asserts len(results); this is both expensive for CI and not a meaningful test (it will pass even if solve rate is 0%). Consider either deleting this test, marking it skipped by default (e.g., @pytest.mark.skipif(not MAKE_VIDEOS, ...) / custom flag), or reducing the seed set and asserting an expected minimum solve rate.

Copilot · 2026-04-01T19:50:54Z

tests/oracles/clutteredstorage2d_medium/test_approach.py

+    shared_env = KinderGeom2DEnv(ENV_ID) if not MAKE_VIDEOS else None
+
+    approach = ClutteredStorage2DOracleApproach(
+        action_space=KinderGeom2DEnv(ENV_ID).action_space,
+        observation_space=KinderGeom2DEnv(ENV_ID).observation_space,


ClutteredStorage2DOracleApproach is initialized with KinderGeom2DEnv(ENV_ID) twice just to access spaces. This repeatedly constructs environments and can slow tests; reuse shared_env (when available) or create a single temp env and pull both action_space and observation_space from it.

Suggested change

shared_env = KinderGeom2DEnv(ENV_ID) if not MAKE_VIDEOS else None

approach = ClutteredStorage2DOracleApproach(

action_space=KinderGeom2DEnv(ENV_ID).action_space,

observation_space=KinderGeom2DEnv(ENV_ID).observation_space,

if MAKE_VIDEOS:

shared_env = None

temp_env = KinderGeom2DEnv(ENV_ID)

try:

action_space = temp_env.action_space

observation_space = temp_env.observation_space

finally:

temp_env.close()

else:

shared_env = KinderGeom2DEnv(ENV_ID)

action_space = shared_env.action_space

observation_space = shared_env.observation_space

approach = ClutteredStorage2DOracleApproach(

action_space=action_space,

observation_space=observation_space,

Copilot · 2026-04-01T19:50:54Z

tests/oracles/clutteredstorage2d_medium/test_approach_debug.py

+)
+
+ENV_ID = "kinder/ClutteredStorage2D-b3-v0"
+DEBUG_SEEDS = [7, 11, 12, 13, 14, 17, 18]


DEBUG_SEEDS is assigned twice; the first list is immediately overwritten, which is confusing and easy to miss when updating seeds. Remove the dead assignment or rename one of them (e.g., keep a commented "candidate" list) so the parametrization reflects the intended seed set.

Suggested change

DEBUG_SEEDS = [7, 11, 12, 13, 14, 17, 18]

Copilot · 2026-04-01T19:50:54Z

tests/oracles/clutteredstorage2d_medium/test_approach_debug.py

+    VIDEO_DIR.mkdir(parents=True, exist_ok=True)
+    ARTIFACT_ROOT.mkdir(parents=True, exist_ok=True)
+    log_path = ARTIFACT_ROOT / f"debug_seed{seed}.log"
+
+    env: Env = RecordVideo(
+        kinder.make(ENV_ID, render_mode="rgb_array"),
+        str(VIDEO_DIR),
+        name_prefix=f"debug_seed{seed}",
+    )


This test unconditionally creates artifact directories and records video (and later asserts an mp4 exists). In CI this can be flaky or fail outright if video encoding deps (e.g., ffmpeg/moviepy) aren’t available, and it also writes files during normal unit test runs. Consider skipping the test unless --make-videos is enabled (like other oracle tests), or at least gate the RecordVideo wrapper + assert video_files behind that flag.

Jaraxxus-Me

Overall LGTM! Left some comments:

Refactor motion planning and reuse what is already in kinder
Split the behavior in to smaller ones, and inside each one we should not have _phase scheduling
Some small requrests on unit tests

Jaraxxus-Me · 2026-04-02T00:44:25Z

src/robocode/oracles/clutteredstorage2d_medium/act_helpers.py

+    return path
+
+
+def plan_holding_base_path(


Thanks for these efforts! But what's the difference between these and BiRRT in the primitives? Can we use BiRRT to get the collision free path instead? https://github.com/tomsilver/robocode/blob/main/src/robocode/primitives/motion_planning.py

Also, for these 2d environments we have existing motion planner here: https://github.com/Princeton-Robot-Planning-and-Learning/kindergarden/blob/90be02c2102de2216092c0b0375c37637b923e37/src/kinder/envs/kinematic2d/utils.py#L273

I think putting this run_motion_planning_for_crv_robot in to https://github.com/tomsilver/robocode/blob/main/src/robocode/primitives/crv_motion_planning.py might be a good idea as we might provide it as a primitive to coding agents later.

So are we going to use the birrt or the crv one? If the latter one should I make it a primitive first ?

Yes, let's use the latter one, and please make it a primitive! I will also use this in my pushpullhook oracle

Qianwei @weiqianwang123, after you implement crv_motion_planning.py, could you also help me implement one more primitive based on this (in another .py file: crv_motion_planning_grasp):

crv_motion_planning_grasp:
So it takes the following inputs: The current object-centric state, the grasping target object, the relative grasping pose, the grasping arm length. Then it outputs a list of (collision-free) waypoints and finally extends the arm to suction the object. The function could also raise two errors: (1) Sucting failed, trying to suction empty space and (2) Sucting failed, no collision free path found.

I guess it will be straightforward to implement this primitive by reusing some of the code already in the oracle behavior? And I hope to try this primitive soon in our next round of experiments! Thanks!

Ok I will do that tonight.

Great! Thank you!

Jaraxxus-Me · 2026-04-02T00:50:47Z

src/robocode/oracles/clutteredstorage2d_medium/behaviors.py

@@ -0,0 +1,1138 @@
+"""Oracle behaviors for ClutteredStorage2D-b3."""


Thanks for these efforts! This must be a lot of work, I would be super surprised if Claude Code can find solution even better than this ORZ.

My only comment is to split this behavior into several smaller behaviors: in side each behavior, we should not have "phase" scheduling, each behavior just have a single initiation condition, a single goal, and a single policy/way point generation.

Ok I will address that

Jaraxxus-Me · 2026-04-02T00:56:19Z

tests/oracles/clutteredstorage2d_medium/test_approach.py

@@ -0,0 +1,131 @@
+"""Tests for the ClutteredStorage2D oracle approach."""


I think we only need something like test_approach_debug.py, I assume you are using this test_approach.py to get results? For that purpose we will have this: https://github.com/tomsilver/robocode/blob/main/experiments/run_experiment.py we will copy the oracle into some load_dir that the agentic approach will be loading from

Jaraxxus-Me · 2026-04-02T00:57:14Z

prpl-mono

Have you changed anything in prpl-mono?

I checked my local git status and git log, and it looks like this is coming from the main GitHub repo side rather than an actual change I made in prpl mono. I did not intentionally modify prpl mono locally.

Okay good, that it should be fine, I updated something in prpl-mono so as long as you use the latest main it is good!

weiqianwang123 added 2 commits March 30, 2026 18:44

initial submit for new oracle for clutteredstorage

fdbdf66

resolve format issues

f3232d3

weiqianwang123 changed the title ~~initial submit for new oracle for clutteredstorage~~ clutteredstorage medium oracle Mar 31, 2026

weiqianwang123 added 15 commits March 30, 2026 21:18

add motion planning,need better pick policy

c120785

pick now works fine,now the problem :1,collision while holding 2,weri…

0e84571

…d stuck in pushing the block in shelf

stuck issue improved,now addressing motion planning while holding blocks

892c708

make some case for stuck issue better (11/20->13/20)

07eae7e

14/20,now addressing motion planning while holding blocks

fbc0f5b

17/20, need to address: 1,stuck in shelf 2,stuck at initial 3,hit

6d910e0

18/20,now addressing :1. stuck in shlef,2. stuck in intial

065fa14

19/20, now address: stuck at initial

6f119d8

76/100,now problem:1,pre insert collision 2,extend arm collision

c240fbe

80/100,now address:1.stuck at initial 2,collision at extending arm

ef0c8ea

88/100.now addressing : collising while extending arm

f105d70

93/100, now adressing : remaining issues

a5d491c

94/100

89e8249

96/100

d6a2f66

address ci checks

d428dc5

weiqianwang123 requested review from Jaraxxus-Me and Copilot April 1, 2026 02:08

Copilot started reviewing on behalf of weiqianwang123 April 1, 2026 19:46 View session

Copilot AI reviewed Apr 1, 2026

View reviewed changes

Jaraxxus-Me requested changes Apr 2, 2026

View reviewed changes

weiqianwang123 added 7 commits April 2, 2026 22:29

95/100, add and use motion planning primitive

50ad5da

add grasping primitive

2a2e5bb

96/100

48b4afd

clear helper file

3e76f04

93/100. change primitive for motion and grasp planning

9788e90

93/100,working for clear behavior

93a46a0

93/100,clean behavior

bb1ff9c

97/100, now addressing some other issues

313771b

-    shared_env = KinderGeom2DEnv(ENV_ID) if not MAKE_VIDEOS else None
-    approach = ClutteredStorage2DOracleApproach(
-        action_space=KinderGeom2DEnv(ENV_ID).action_space,
-        observation_space=KinderGeom2DEnv(ENV_ID).observation_space,
+    if MAKE_VIDEOS:
+        shared_env = None
+        temp_env = KinderGeom2DEnv(ENV_ID)
+        try:
+            action_space = temp_env.action_space
+            observation_space = temp_env.observation_space
+        finally:
+            temp_env.close()
+    else:
+        shared_env = KinderGeom2DEnv(ENV_ID)
+        action_space = shared_env.action_space
+        observation_space = shared_env.observation_space
+    approach = ClutteredStorage2DOracleApproach(
+        action_space=action_space,
+        observation_space=observation_space,

		@@ -0,0 +1,1138 @@
		"""Oracle behaviors for ClutteredStorage2D-b3."""

		@@ -0,0 +1,131 @@
		"""Tests for the ClutteredStorage2D oracle approach."""

Conversation

weiqianwang123 commented Mar 30, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 1, 2026

Choose a reason for hiding this comment

Uh oh!

Jaraxxus-Me left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants