pybullet_domino+agent_sdk: certificate criterion experiments, BeliefProbe rename, dense sigma grid (sysID stack 5/8) - #100
Merged
Conversation
…set rule A target rammed off the table upright acquires its roll only in freefall and lands flat on the floor; the roll-only Toppled classifier scored that as a solve (caught in run_20260723_110212's explore replay at mu=0.1, where the low-friction relay slides like a puck and shoves purple off the edge still standing). The Toppled predicate deliberately stays roll-only so a topple that happens on the table and then slides off still counts at the final state. The loophole is closed trajectory-side instead: on goal-reaching episodes the certificate now requires every goal target's fall to have BEGUN at table level (z at its first fallen-threshold crossing; new _fall_index helper, _topple_onset refactored onto it). On-table crossings pivot on the base edge at z 0.42-0.47 while an off-table shove crosses in freefall at z 0.07, so the band separates cleanly. DominoComponent.toppled_on_table (roll + table level) survives as the design-time check in both min-block generation probes, and the goal texts state the enforced rule. Also: the two test classes in test_pybullet_domino_composed.py had __init__ constructors, so pytest silently never collected them; fixing that exposed a stale grid test and a real grid_component bug (extract_feature crashed on grid-index loc names instead of the documented None fall-through; it now prefers the stored coordinate attributes and parses names defensively). Note on task caches: any .py edit under the domino packages changes _domino_code_digest, so min-block tasks regenerate and same-seed tasks can differ from old runs; replays of old runs must load the run's cached task JSON (regeneration is not byte-stable).
…be.py) The agent-facing sim handle is a stateful probe over the belief simulator (BeliefSim), not an env; name the class for what it probes. Mechanical rename across source, prompt text, and tests; the explore_python namespace now binds BeliefProbe for extra instances. Probe result/exception types and build_probe_namespace keep their names (probe stays the verb).
…topple criterion The on-table topple requirement (81cbf0b) broke min-block TEST task generation: the turn tasks' differentiation filter probes each leg at the planning friction (0.1), where the rammed leg target slides off the probe table - with on-table required, no k ever counted as toppling, so believed K* was None for every candidate, the filter dropped them all, and the 76-attempt cap shipped empty caches ("generates 0 test tasks", launches 13:53-14:09 on 2026-07-23). New _probe_toppled dispatches on whose physics is loaded: at the TRUE friction the probe stands in for the certificate (topple must happen on the table); at the believed/planning friction it stands in for the uncalibrated planner, whose success check is the roll-only Toppled goal atom - an off-table shove convinces that planner, and believed-K* must measure what the planner believes, not what would certify. Probe memo keys already include the physical override, so the two criteria never share entries. Verified standalone: seeds 0-2 each generate 1/1 turn test task (k*=2) again; the broken window's empty caches are orphaned by the code-digest change.
…ll-only topple criterion" This reverts commit b1f5bc4.
…icate onset rule" This reverts commit 81cbf0b.
The all-train-attempts-solved criterion certifies the learned model, but it could fire on cycle 0 before any learning ran (explorer successes come from the initial mental model), skipping learning for the whole run (e.g. run_20260723_141019 tested an untrained model). Gate early stopping on model_has_learned: offline demos, a loaded approach, or a prior online learning update.
…eview The capture gate's +-1-sigma margin check was two endpoint rollouts; near a feasibility boundary success is speckled in the physical params, and run_20260724_140531's capture passed both endpoints while failing deterministically at the true friction between them. physics_sigma_points now returns an even fit-space grid (agent_plan_validation_physics_margin_points, default 32: empirically a 16-point grid missed that capture's failure hole, 32 catches it), and the agent can pre-check its exact plan with sim.run(plan, physics_sweep=True) - one deterministic rollout per gate point - instead of discovering PARAM-SENSITIVE one submission at a time. Also refreshes the stale heavy-block settings comment (heavy generation verified at the default true friction 0.5).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fifth of eight PRs landing the domino-sysid branch on master in sequence (after #96, #97, #98, #99). Tree is identical to the corresponding slice of that branch rebased onto current master.
This chunk
Two certificate-criterion experiments and their deliberate reverts are kept as history (the off-table loophole close broke believed-K* test-task generation; both were backed out on purpose). Net changes: the BeliefProbe rename, the model-has-learned eligibility gate so cycle-0 can't early-stop before any learning has happened, and the 32-point sigma grid with the agent-side physics_sweep preview (16 points empirically missed the k*-boundary failure hole at true friction; 32 catches it).
🤖 Generated with Claude Code