Physics Studio Contract Verification¶

Current status¶

Studio contract: VERIFIED

This verification establishes the implementation boundary: Studio uses the author training/player paths, changes only the articulated task boundary, and publishes one common rollout/evaluation/rendering interface. It deliberately does not claim that five iterations establish learned task quality, that an old dependency stack produces identical historical metrics, or that an author repository with a missing success implementation can supply an official success number.

Method	Studio-contract evidence	Not claimed here
PHC-X	released author checkpoint/player, fresh 150-frame common rollout, shared evaluation and H.264 replay	training quality (PHC-X is tracking-only)
CoDA	delivered author ZIP audited; 5 complete author PPO iterations at 4096 envs (655,360 transitions, 240 updates), fresh author checkpoint reload, full-\(J\) common rollout and shared replay	final learned success rate; the delivered author `MimicCriterion.compute_success` raises `NotImplementedError` on completion
InterMimic	5 complete author PPO iterations at the author `numEnvs: 4096` (655,360 transitions, 240 updates), fresh reload, full-\(J\) common rollout and replay; unmodified `eval_teacher.sh` completed at official `--num_envs 1024` for 27/27 sequences	final learned success rate or cross-version metric parity
RePHO	current-q0 author workers/player write a full-\(J\) common rollout and shared replay	training verification and the long validation/repair protocol, explicitly deferred
ours	outside this author-code fidelity audit; its Studio integration evidence is tracked separately	this page does not judge `ours` quality or freeze its paper version
HDMI-SMPL-X	not part of this verification	separate adaptation review is required

Evidence currently on disk includes:

output/physics_studio_verification/final_author_20260729/phcx_author_eval/eval/
output/physics_studio_verification/final_author_20260729/coda_4096/eval/
output/physics_studio_verification/final_author_20260729/intermimic_4096/eval/
output/physics_studio_verification/final_author_20260729/repho_512_current/eval/

All four published evaluation directories contain only humanoid.xml, common_rollout.npz, evaluation.json, replay.mp4, and run_record.json. The InterMimic author training configuration is numEnvs: 4096. Its official quantitative evaluation script explicitly overrides to --num_envs 1024; the completed native run covered 27/27 sequences and reported 0/27 success. This is recorded as an author-entrypoint execution result, not baseline-quality or cross-version-fidelity evidence. The earlier 1024 Studio training run remains a disclosed resource pilot; the current 4096-env pilot is the native-capacity format verification.

The formal InterMimic experiment preserves the author interaction budget of 1,024 surface points per future horizon. Its object observation is replaced as one semantic unit: 21-D rigid-root tracking becomes 21-D active-link pose/velocity tracking, and the 219-D rigid-mesh interaction graph becomes link-FK surface interaction. AL samples the complete active-link surface; Full samples the union of all link surfaces, always keeping 1,024 points. The human observation, two future horizons, input size, action, teacher, author PPO/GAE/buffer/curriculum, seed and budget stay identical. All joints remain physically free; the active link only identifies the supervised task target. A bridge that retains rigid-root 21-D tracking while substituting only some contact points into the interaction graph is diagnostic only, not a reportable InterMimic row.

Old outputs under output/runs/physics/studio_* used earlier generic training loops. They are interface/debug results only and must not be reported as baseline performance.

Current same-input evidence is deliberately kept outside the runtime:

scripts/physics/verification/coda_vendor_snapshot.py verifies the delivered ZIP SHA-256 and every one of its 605 files. The only permitted ZIP modification is parse_task.py; the only additions are the four explicit physhoi/studio/* files;
scripts/physics/verification/author_source_digest.py checks the unchanged PHC-X, InterMimic, and RePHO PPO collection/update/GAE functions against their pinned upstream revisions;
tests/physics/test_author_numeric_seams.py checks the released InterMimic and RePHO checkpoints in clean processes: strict state-dict reload, external normalizer, and a fixed-observation deterministic action/value fingerprint. It also checks the existing external-normalizer, policy distribution, and one RePHO PPO-update seam on identical synthetic tensors. Additional component fixtures are useful release evidence, not a Studio contract gate.

These checks establish algorithm preservation at the boundary; they do not turn a five-iteration pilot into a training-quality or task-success result.

Question being verified¶

For a fixed canonical HOI reference and one shared full-\(J\) articulated scene, do PHC-X, CoDA, InterMimic, and later eligible methods:

execute their complete author algorithm;
differ only through their method logic and minimal articulated adaptation;
produce comparable Success/Fail and visualization from the same common output boundary?

The primary result is case-level binary Success/Fail. Diagnostics include per-phase progress, contact timing, early termination, object-root drift, and unsupervised-joint drift. Unsupervised-joint drift is never a binary failure condition by itself.

Experimental separation¶

Official reproduction and Studio benchmark are tested independently.

Official reproduction¶

run the pinned author entrypoint, native input, asset, config, and checkpoint;
do not enable a Studio task or import the main pipeline;
retain author logs and videos only as reproduction evidence;
do not aggregate these results with Studio results.

Studio benchmark¶

load the fixed case.json + result.pt pair and contact labels;
use the shared full-\(J\) articulated scene and case-specific human;
start the author's complete trainer, player, or coordinator;
allow only explicit articulated task/manager/recorder additions;
write the common rollout directly;
run the shared binary evaluator and common renderer.

An official result cannot substitute for a Studio result, and a Studio checkpoint cannot silently substitute for a released author checkpoint.

Fixed Studio boundary¶

All methods receive:

the same canonical case and dynamic reference;
the same contact-reference labels;
the same source-to-control timeline rule;
the same full object topology, free joints, limits, mass, collision, friction, support plane, and static boxes;
the same 52-body public human order;
the same evaluation thresholds.

All object joints remain physically free. Supervised joint names select the task objective only. CoDA currently accepts one supervised reward target, but its environment must still simulate and record all other object joints.

Each method may retain author-native:

control and simulator frequency;
observation and action representation;
reward formula and weights;
reset distribution and curriculum;
environment count, horizon, minibatch, mini-epochs, and optimizer;
checkpoint layout and native counters.

These are method constants, not quantities that Studio normalizes. Timeline conversion occurs before the author task, and any author coordinate canonicalization is inverted before common recording.

Allowed implementation changes¶

The root backend may:

convert canonical data to a temporary native input;
prepare author config values that point to the shared scene;
launch the author entrypoint;
validate required native outputs;
publish the common outputs.

An opt-in submodule patch may:

register an articulated Task or manager;
load the shared object/support actors;
expose method-required object/contact observations;
add the minimal articulated reward term;
reset all object states;
record full-\(J\) state and contact telemetry.

The shared runtime uses rl-games 1.6.1. A mechanical compatibility patch for that upstream API is allowed when it only preserves the original operation under the new API (for example an import, equivalent argument spelling, or container access). It must preserve input tensors, PPO schedule, buffer fields, checkpoint layout, and player behavior. The final diff audit and targeted source/fingerprint tests check that narrow boundary; it is not a license to replace any author algorithm.

It may not replace or simplify:

author network and normalizer;
trajectory collection and buffer;
return/advantage calculation;
PPO update, minibatches, optimizer, or scheduler;
curriculum, RSI, or termination unrelated to the object adaptation;
checkpoint save/restore;
author player or RePHO coordinator.

No baseline may use a trainer defined in pipeline/physics.

Public and temporary files¶

Public evaluation output is:

eval/
  humanoid.xml
  common_rollout.npz
  evaluation.json
  replay.mp4
  run_record.json

Training preserves the author's checkpoint and TensorBoard directories. run_record.json stores direct facts only: method, mode, case, seed, input and output paths, checkpoint paths, and counters.

CoDA's 62-node reference, native human XML, generated author config, and other backend bridge files stay in a temporary directory. InterMimic/RePHO packed motion inputs and PHC-X native recorder output are temporary for the same reason. Evaluation and rendering never consume them.

Source inspection, captured-tensor checks, checkpoint inspection, hashes, and multi-run statistics belong to scripts/physics/verification/; they are not runtime fields or runtime dependencies.

Studio verification gates¶

The gates below deliberately verify the implementable Studio contract. They do not make a result-quality claim.

Gate 1: one strict common boundary — passed¶

There is one load_physics_input(case_json, result_pt) path, no runtime fallback/legacy readers, one common-rollout validator, and one renderer. The same strict tests cover finite values, shapes, timelines, source conversion, and public 52-body ordering.

Gate 2: full-\(J\) physical scene and reset — passed¶

Scene preparation, reset, recorder, evaluator, and renderer preserve every free object joint in order. Final stateInit=Start evaluation resets to the canonical reference's q(0) with zero velocity. Author stateInit=Hybrid/Random training instead resets the object with the sampled human at the same reference's absolute q(t); it never uses a relative joint offset. A shared-Isaac-asset PhysX perturbation test shows an unsupervised joint has no drive and moves physically. Its drift is recorded only as a diagnostic.

Gate 3: author-algorithm boundary — passed¶

The Studio runtime contains no generic policy or trainer. PHC-X launches the author player; CoDA and InterMimic launch their author trainer, buffer, GAE/PPO, checkpoint, and player; RePHO launches its author workers/player. The final submodule diff is confined to explicit Studio articulated task, object/contact/reward/reset/recorder code plus the narrow rl-games-1.6 sentinel compatibility seam. It does not replace author network, trainer, buffer, optimizer, curriculum, checkpoint, player, or coordinator.

Gate 4: execution witness — passed within declared scope¶

PHC-X completed a fresh author-player rollout. CoDA and InterMimic each completed five full native outer iterations at their author 4096-environment capacity, saved a checkpoint, then reloaded it in a fresh author player. RePHO is functional-interface verified through the current-q0 author workers and player; its formal training/validation branch is intentionally deferred.

Gate 5: common-output and visual sanity — passed¶

Every published rollout strictly loads with pickle disabled, contains full-J state, is evaluated solely by the shared evaluator, and is rendered solely from the common rollout plus static case assets. Frame-0 and selected keyframes were visually inspected for body order, object/support alignment, active-link color, camera, and reset-frame artifacts. The final output directories listed above are the reproducible witnesses.

The retained verification scripts add targeted evidence only: source-body digests for untouched author PPO paths, released-checkpoint normalizer/action fingerprints, CoDA ZIP audit, analytic FK/order checks, and the independent passive-joint PhysX test. They are not runtime dependencies.

Separate, non-blocking evidence¶

The following remain valuable before reporting a result table, but cannot block the Studio implementation because they are either task-quality questions or unavailable in the delivered author code:

long native training and held-out success rates for each baseline;
an archived author environment's historical metric signature;
CoDA's native strict-success result while the delivered MimicCriterion.compute_success raises NotImplementedError;
RePHO's long validation/repair protocol, deferred by experiment scope.

Method-specific notes¶

PHC-X¶

PHC-X is evaluation-only. The author checkpoint, player, PNN, normalizer, and action path must remain intact. No PHC-X training claim is made.

The previous full20 b004-0017 record with phc_success_rate=1.0 used kinematic_reference: it establishes only historical human tracking while the object followed its reference. Its overall physics result was still Fail. The Studio path uses passive-dynamic full-joint physics and does not accept that legacy mode as an input fallback.

CoDA¶

CoDA must use its author task, model, buffer, PPO implementation, scheduler, checkpoint, and player. The one-target articulated reward is retained. The adapter may construct its native 62-node reference, but all object joints remain free and the final output is the public 52-body common rollout.

InterMimic¶

InterMimic must run InterMimicAgent with the author buffer, return/advantage calculation, PPO update, curriculum, checkpoint, and player. The Studio task adds only the articulated scene, a matched active-link object observation and reward, link-FK interaction-graph surfaces, and recorder logic. Main-paper rows are InterMimic-AL zero-shot, InterMimic-AL + 2k PPO, and InterMimic-Full + 2k PPO; their exact protocol is frozen in physics_baselines.md. InterMimic-AL scratch is supplement-only, triggered only by a reviewer request and run with a pre-registered convergent author budget rather than the 2k transfer budget. The fixed 21-D author object slot supports one supervised active link per run; multi-target cases fail explicitly at this adapter boundary while the shared scene remains full-J.

RePHO¶

RePHO functionality has been exercised through its current-q0 author workers and player. The user has explicitly deferred long RePHO training and the validation/repair protocol. Its status here is therefore functional-interface verification only, never training quality or a benchmark comparison.

Completion rule¶

Studio contract: VERIFIED means Gates 1–5 pass for the methods and scope stated in the table, current code has been reviewed independently, and official reproduction remains separate from Studio outputs. It means the algorithm boundary and common physical/output contract are ready for real baseline training; it does not imply a learned-success claim.

Before publishing a result table, record the actual training budget, seed set, held-out Success/Fail rates, and any available official-reproduction evidence. The deferred RePHO protocol remains a format/interface check until it is run.