PaperlayerView exact revision on arXiv

Source-linked AI summary

EMPIRIC: Experiment-Driven Learning of Residual World Models for Robot Planning

Yichao Liang, Amber Li, Dat Nguyen, Emily Bunnapradist, Michelangelo Naim, Sreela Kodali, Matteo Merler, Bowen Li, Kiran Gopinathan, Yiyun Liu, Nikhil Pimpalkhare, Joshua B. Tenenbaum, Adrian Weller, Zenna Tavares, Tom Silver, Kevin Ellis

arXiv:2609.35047v1cs.ROcs.AIcs.LG

TL;DR

Robots need to learn mechanisms that a supplied physics engine does not represent, while accounting for uncertainty in parameters and hidden state. EMPIRIC writes those mechanisms as code and uses the extended simulator to plan and select experiments. Across five simulated domains it solves more runs than the baselines, and it also learns enough on a physical robot to solve a manipulation task.

  • Problem

    A physics engine may omit unfamiliar mechanisms, while robots must infer their parameters and hidden state from noisy, partial observations to plan safely and choose informative experiments.

    Source: §1 · p. 1

  • Method

    EMPIRIC extends a base physics simulator with executable mechanisms and recurrent hidden state, then infers parameters and state to plan, experiment, and revise its model.

    Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 5

  • Results

    Across five simulated domains, EMPIRIC solves all 25 runs versus 16 for Direct agent, 14 for Direct + scene, and 16 for Standalone sim, while using fewer environment steps than the baselines overall.

    Source: §4.1 · p. 7, §6 · p. 9

  • Takeaways & Limitations

    On a physical robot, EMPIRIC learns wind forces and domino masses well enough to solve a manipulation task; the results also suggest explicit uncertainty matters most for irreversible actions.

    Source: §6 · p. 9

    Experiments assume supplied scene geometry and predefined object features, although object poses are noisy and mechanism state is hidden.

    Source: §6 · p. 10

Abstract

from arXiv · show

A robot should be able to learn through experiments how unfamiliar objects behave and interact, then plan with that knowledge. It need not start from scratch: physics engines supply knowledge of motion and contact, but can omit entire mechanisms, such as glue curing, water heating, or wind. We present EMPIRIC, an agent that learns a residual world model: a physics engine extended with code for the missing mechanisms. The learned programs can introduce new forces, constraints, and hidden state, and Bayesian inference estimates their parameters and states from noisy observations. The resulting model lets the agent predict the outcomes of actions, choose informative experiments, and revise its hypotheses when predictions fail. Across five simulated domains, EMPIRIC learns interpretable, reusable models, and solves more tasks with fewer environment interactions than all three baselines. On a physical robot, it learns wind forces and domino masses to solve a manipulation task. Website and code: https://yichao-liang.github.io/empiric

1 INTRODUCTION

EMPIRIC addresses how robots can learn physical mechanisms missing from a generic physics engine, then plan with the resulting model under uncertainty. It extends the engine with executable code and hidden state, and is evaluated in simulation and on a physical robot.

1 INTRODUCTION

  • A physics engine can predict rigid-body motion yet omit mechanisms such as airflow or glue curing, which geometry reconstruction alone does not identify.The robot must infer parameters and potentially unobserved evolving state from few, noisy, partial observations.Source: §1 · p. 1
  • EMPIRIC extends a generic rigid-body simulator with executable programs for missing mechanisms, including forces, constraints, and recurrent hidden state.For example, glue code can accumulate contact time and create a bond after a curing threshold.Source: §1 · p. 2
  • The agent maintains beliefs over mechanism parameters and hidden state from noisy observations, using them to choose plans or experiments and revising code when predictions fail.Uncertainty matters because it affects plan safety and which experiment is informative.Source: §1 · pp. 1–2
  • Across five simulated domains, EMPIRIC solves all evaluated runs with fewer environment steps than baselines, and is also demonstrated in one physical domain.The evaluation uses a continual protocol of training tasks followed by test tasks.Source: §1 · p. 3

2 PARTIALLY OBSERVABLE CONTINUAL MANIPULATION

The paper frames continual manipulation as solving sequential tasks with an unknown transition function and noisy observations. Runs combine training and test tasks under a shared interaction budget and are scored on completing every task and environment steps.

2 PARTIALLY OBSERVABLE CONTINUAL MANIPULATION

  • The environment has an unknown transition function, and observations provide noisy object features and images while omitting quantities such as glue cure progress.The agent receives parameterized skills with continuous parameters and closed-loop controllers.Source: §2 · p. 3
  • Each task specifies an initial state, a natural-language goal, and a binary reward over observation-action trajectories that can constrain how the goal is reached.The evaluator checks the noise-free true trajectory, including task constraints.Source: §2 · p. 3
  • A run has training tasks followed by test tasks, shares one environment-step budget across them, and is scored on solving every task and the steps used.The agent may retain experience and continue learning across tasks; an unsolved test episode can end the run early.Source: §2 · p. 3

3 METHOD

EMPIRIC extends a supplied physics simulator with learned mechanisms, estimates uncertain parameters and state, then uses simulation to plan, seek information, monitor execution, and revise its model.

3 METHOD

  • EMPIRIC edits simulator code, infers parameters from recorded interactions, and simulates plans before execution, with its sandbox persisting across tasks.It can gather observations or revise its model after a failed simulation.Source: §3 · p. 3

3.1 RESIDUAL WORLD MODELS

  • The residual program augments base dynamics with mechanisms and recurrent state, allowing the model to represent forces, constraints, and quantities such as water volume.When observations cannot be explained by any parameter setting, the agent can add a mechanism, such as contact-time accumulation followed by a glue bond.Source: §3.1 · pp. 3–4
  • The learned simulator is deterministic; probabilistic dynamics are left to future work, though the formalism could represent them through hidden parameters.Source: §3.1 · p. 4

3.2 PARAMETER AND STATE INFERENCE

  • The agent approximates a belief over parameters and base state from replay data and episode history, tying each parameter draw to the recurrent state implied by that draw.The parameter posterior uses a mean-field approximation, while the base-state posterior conditions on observations under declared sensor noise.Source: §3.2 · p. 5

3.3 PLANNING AND INFORMATION SEEKING

  • The agent rehearses skill sequences across joint belief draws to estimate task success, and ranks candidate experiments by expected information from noisy predicate readings.It decides whether the expected information is worth the environment cost.Source: §3.3 · pp. 5–6

3.4 EXECUTION MONITORING AND MODEL REVISION

  • After each skill, the agent checks expected predicates; a failed skill or unmet predicate stops the sequence and can prompt replanning, information gathering, belief updates, or code edits.Every environment step counts toward the budget, including experiments.Source: §3.4 · p. 6

4 EXPERIMENTS

The experiments compare EMPIRIC with alternative agents and ablations across five simulated manipulation domains, then test it on a physical robot. EMPIRIC achieves higher overall success than the main baselines, while the results also examine interaction costs, uncertainty, and real-robot planning.

4.1 SIMULATED EXPERIMENTS

  • Five simulated tabletop domains test domino cascades, hidden glue bonds, balloon lift, water filling and heating, and fan-driven ball stopping, using noisy observations and five seeds.Each run requires solving its training and test tasks; Balloons uses two training tasks.Source: §4.1 · p. 6
  • All 25 runs were solved by EMPIRIC, compared with 16 for Direct agent, 14 for Direct + scene, and 16 for Standalone sim.; it used fewer steps than baselines in most domains.In every domain, EMPIRIC reached a 60% solve rate with fewer steps than any baseline that reached it.Source: Figure 3 · p. 7, §4.1 · p. 7
  • Baseline failures often extrapolate beyond observations, while EMPIRIC’s residual model reuses engine physics and robot skills to generalize to new arrangements.The learned programs remain useful for configurations such as turning cascades and longer glued beams.Source: §4.1 · p. 7
  • EMPIRIC matches Oracle dynamics at 25/25 successes, using 12–18% more steps in Domino, Bridge, and Balloons, 74% more in Boil, but 8% fewer in Fan.Oracle receives ground-truth dynamics but still estimates state from noisy observations and chooses its actions.Source: §4.1 · p. 7
  • In Bridge, the model mispredicts the lifted beam’s pose, yet the agent screens placements and measures the assembled beam’s tilt to choose its final release height.This illustrates that a learned model can still support decisions when its predictions are imperfect.Source: §4.1 · p. 8
  • Removing harness fitting reduces success to 20/25 runs, while removing explicit uncertainty support reduces it to 21/25.Some ablated agents independently rebuilt parts of the omitted capabilities, and in Boil or Fan sometimes used fewer steps than EMPIRIC.Source: §4.1 · p. 8
  • In Balloons, EMPIRIC solves 5/5 runs versus 2/5 without harness fitting and 3/5 without explicit uncertainty; all five ablation failures are ceiling bursts.The releases are irreversible, and EMPIRIC uses prior peak-prediction errors to reject a release predicted to leave only 15 mm of ceiling clearance.Source: §4.1 · p. 8

4.2 REAL-WORLD APPLICABILITY

  • On the robot, a 10.1±2.3 cm predicted slide matches the measured 9.9 cm after two gusts let EMPIRIC infer wind decay and both domino masses.For the test task, it stands the gray domino upwind so its fall knocks the green domino into the target patch.Source: §4.2 · pp. 8–9

5 RELATED WORK

The related work spans executable world models, hybrid dynamics learning, planning abstractions, scene reconstruction, and coding agents. EMPIRIC is distinguished by learning omitted physical mechanisms through interaction while retaining a physics engine.

Code world models

  • Code-world-model approaches learn executable transition programs, partially observable models, probabilistic experts, or models from text and video; EMPIRIC instead learns missing mechanisms from noisy interactions while retaining engine physics.The cited approaches include WorldCoder and TheoryCoder, GIF-MCTS (Dainese et al., 2024), and PoE-World.Source: §5 · p. 9

Learning physical dynamics

  • Hybrid simulators, parameter inference, active identification, and uncertainty-aware control address parts of physical dynamics learning; EMPIRIC also revises its mechanism program rather than fitting only a fixed model family.The passage cites hybrid models (Heiden et al., 2021), simulation-based inference (Zhu et al., 2025), and active identification.Source: §5 · p. 9

Abstractions and processes for planning

  • Planning abstractions learn operators, predicates, skills, or causal processes; EMPIRIC complements them with low-level mechanisms that interact with engine physics to determine skill outcomes.Examples include NSRTs, VisualPredicator and pix2pred (Athalye et al., 2026), and ExoPredicator.Source: §5 · p. 9

Scene reconstruction and policy transfer

  • Scene reconstruction and policy-transfer work supplies or aligns scenes and interactions, whereas EMPIRIC takes a scene as input and learns omitted mechanisms such as wind or heating from new interactions.The passage cites scene reconstruction and digital-twin policy transfer (Torne Villasevil et al., 2024; Han et al., 2026; Qureshi et al., 2025).Source: §5 · p. 9

Coding agents for control

  • Coding agents can write control programs or improve policies through feedback; EMPIRIC continually revises a forward simulator and infers its parameters alongside its action programs.The passage contrasts this coupling with simulator-probe approaches that synthesize policies frozen for evaluation, citing AgenticGenPlan.Source: §5 · p. 9

6 CONCLUSION AND LIMITATIONS

The conclusion presents EMPIRIC as a way to learn unfamiliar mechanisms by extending a physics simulator, planning with the resulting model, and revising it after prediction failures. Results cover five simulated domains and a physical-robot task, while the agent’s inputs and computational cost constrain the evaluation.

6 CONCLUSION AND LIMITATIONS

  • Across five simulated manipulation domains, EMPIRIC solves every evaluated run and outperforms baselines in overall success and sample efficiency; on a physical robot, it learns enough from two gusts for a two-domino cascade.The authors suggest the world model generalizes to new configurations and that explicit uncertainty matters most for irreversible actions.Source: §6 · p. 9
  • The experiments provide scene geometry and predefined object features; the agent does not reconstruct scenes or learn image-based features, and its interaction savings carry higher computational cost.Median runs use 424 simulator rollouts and 162 language-model turns, with a 107-minute runtime, versus 115 turns and 64 minutes for Direct agent.Source: §6 · p. 10

A DOMAIN AND EVALUATION DETAILS

The appendix specifies the observation and interaction setup, supplied robot skills, and what counts toward the interaction budget. It also describes the information agents may retain across tasks.

A.1 OBSERVATIONS, FEEDBACK, AND INTERACTION

  • Visible numerical features are noisy observations of state, while discrete features and robot proprioception remain exact; hidden quantities such as glue cure progress are unobserved.Noise is drawn once per environment step, so repeated observations without stepping provide no independent measurements.Source: §A.1 · p. 14
  • Agents receive named, typed object features and images, and can use supplied closed-loop skills with object arguments and continuous parameters or issue primitive actions.The skills include Pick, Place, Push, MoveTo, and Wait.Source: §A.1 · p. 14
  • Every primitive step, including waiting, counts toward the interaction budget; agents may retain recordings, programs, and journal notes across tasks and continue learning during testing.Training permits resets of the current task, whereas testing does not.Source: §A.1 · p. 15

A.2 DOMAINS AND OBSERVATION SCALES

The simulated domains test manipulation across contact, structural assembly, irreversible balloon release, liquid heating, and airflow. Observation noise is applied to feature measurements, while rendered images are noise-free.

A.2 DOMAINS AND OBSERVATION SCALES

  • Domino tests a high-friction cascade that must turn while satisfying contact constraints.Source: §A.2 · p. 15
  • Bridge tests learning to join and manipulate blocks, with a four-block span seated on two supports at test.Source: §A.2 · p. 15
  • Balloons tests payload lift and settling within a target band, with irreversible releases and ceiling contact that can burst balloons.Test solutions free at least two balloons whose colors never shared a training rack; releasing the weakest first bursts one on the ceiling.Source: §A.2 · p. 15
  • Boil trains with one jug and tests two-jug filling, transfer, and heating under a spill constraint.Source: §A.2 · p. 15
  • Fan tests airflow control to move a ball across platforms and a downhill ramp, then stop it at a target.Each domain has one training and one test task except Balloons, which has two training tasks.Source: §A.2 · p. 15
  • Feature observations are noisy, rendered images are noise-free, and scalar noise is additive and unclipped.Position noise is measured in millimeters, orientation noise in radians, and scalar noise in each feature’s units.Source: §A.2 · p. 15, Table 1 · p. 15

A.3 BUDGETS, SUCCESS, AND METRICS

The appendix specifies run budgets, success and step metrics, and the provenance of comparisons and recorded visualizations. Task and step observations within a run are not independent replicates.

A.3 BUDGETS, SUCCESS, AND METRICS

  • Pooled step caps are 10,000 for Domino, Boil, and Fan, 20,000 for Bridge, and 15,000 for Balloons, with a 48-hour wall-clock guard.There is no separate episode step horizon; failed runs can end early, and remaining tasks count as unsolved.Source: §A.3 · p. 15
  • A run’s whole-run success counts solved training and test tasks, while its step cost totals environment steps including resets.Figure 3 reports these measures across five seeds in each domain-agent cell.Source: §A.3 · p. 15
  • Mean steps include failures, so a low average alone cannot establish efficiency because failed runs may end early.Source: §A.3 · p. 16

A.4 COMPARISON SCOPE AND PROVENANCE

  • The seven agents share Claude Opus 5, task interfaces, a skill library, and interaction budgets, but receive different modeling support.No explicit uncertainty removes noise scales, state smoothing, and parameter-uncertainty support together, so it does not isolate any one component.Source: §A.4 · p. 16
  • For every reported run, the archive records its seed, source run, code revision, configuration, termination reason, and task outcomes; all reported runs finished.Seeds of one agent can use different tools and learn different programs.Source: §A.4 · p. 16
  • Figure 5 traces experiments, learning, learned-model checks, and test-task completion across Domino, Boil, and Fan.Its dashed frames replay checked plans using the simulator, parameters, observations, and code version available to the agent at the time.Source: §A.4 · p. 16, Figure 5 · p. 17

B IMPLEMENTATION DETAILS

EMPIRIC’s implementation combines an editable simulator subclass and recurrent state with fitted parameter and state beliefs. The agent uses rehearsals and monitored interaction to plan, gather information, and revise its model, while the resulting belief remains approximate.

B.1 AGENT WORKSPACE AND TOOLS

  • The sandbox stores simulator code, predicates, trajectories, and a journal, and provides tools to fit parameters, inspect errors, refine skills, and rehearse plans.Program edits stale the previous fit, so the agent is instructed to refit before planning with revised code.Source: §B.1 · p. 18

B.2 SIMULATOR SUBCLASS AND RECURRENT STATE

  • The simulator subclass retains base physics and adds a hook for missing mechanisms, which can apply forces, constraints, or simulated-quantity updates.Its recurrent state represents the model’s hypothesis about hidden physical state.Source: §B.2 · pp. 18–19
  • The recurrent-state update uses sanitized observations, preceding actions, and parameters, and runs once per transition during both simulation and observed execution.Planning branches keep independent state copies; repeated observations without an environment step do not advance the state.Source: §B.2 · pp. 18–19

B.3 BELIEF CONSTRUCTION

  • For a fixed program, fitting pools recorded episodes using a replay loss that emphasizes settled states and downweights very large mismatches.A generalized posterior temperature is raised when minimum replay loss exceeds its nominal scale, widening the parameter belief.Source: §B.3 · p. 20
  • Parameter uncertainty is approximated with independent coordinate factors, while object-state beliefs average recent still-frame observations and account for uncertain motion-change times.Independent factors can understate the spread of correlated parameters; recent-observation state inference avoids feeding potentially erroneous learned dynamics back into state estimates.Source: §B.3 · pp. 20–21

B.4 REHEARSAL, EXECUTION, AND MONITORING

  • Plan rehearsals evaluate skill sequences across joint parameter and state draws, report simulated outcomes, and refine unbound skill parameters against expected predicates.The rehearsal success estimate does not anticipate replanning after later observations.Source: §B.4 · p. 21
  • The agent ranks candidate experiments by predicate information, then executes its highest-estimated plan or gathers information when it judges confidence insufficient.Monitoring can stop execution after a failed skill or unmet predicate, returning control to the agent to revise the plan or model.Source: §B.4 · pp. 21–22

B.5 APPROXIMATIONS

  • The belief is not calibrated by construction, and the reported runs include no held-out-history calibration tests.Task success does not substitute for calibration evaluation.Source: §B.5 · p. 22

C REAL ROBOT EXPERIMENTS

The physical-robot appendix describes EMPIRIC’s experiment-driven modeling loop and identifies two components needed for deployment: an observation model and reliable manipulation skills.

C REAL ROBOT EXPERIMENTS

C.1 PHYSICAL WORKSPACE AND EVALUATION ENVIRONMENT

The physical evaluation uses a fixed fan-and-button setup with two dominoes, while perception reconstructs scenes for simulator fitting and planning. The robot uses calibrated skills to interact with these objects and perform successive experiments.

C.1 PHYSICAL WORKSPACE AND EVALUATION ENVIRONMENT

  • The fixed fan and button control two dominoes of equal nominal dimensions but different materials; their masses and material properties are withheld from the agent.The agent also does not receive the fan motor’s supply voltage or current; the fan and button remain fixed while it rearranges the blocks.Source: §C.1 · p. 22

C.2 PERCEPTION AND SIMULATION-BASED INFERENCE

  • Scene reconstruction uses SAM 2 masks and stereo depth, and replayed actions render predicted masks and depths to fit base and residual simulator parameters.The fitted simulator predicts state evolution from the latest perceived object poses and measured robot joints, which initialize planning.Source: §C.2 · pp. 22–23

C.3 PLANNING AND SKILL EXECUTION

  • Physical interaction uses motion planning, force-controlled grasps, recovery skills, and a calibrated button press triggered by estimated contact force.The press trigger is 5.21 N, 1 N above the measured actuation force; the skill holds the button down after triggering.Source: §C.3 · pp. 23–24

C.4 FAN AND DOMINOES WITH DIFFERENT MASSES

Two single-block probes support a learned wind model and a successful domino cascade into a relocated target; the recorded demonstration is one run on one bench.

C.4 FAN AND DOMINOES WITH DIFFERENT MASSES

  • The agent’s residual wind program applies constant force during the button hold and exponentially decaying force after release.The program’s parameters include wind-force magnitude and decay time.Source: §C.4 · p. 24, Table 2 · p. 26
  • Green-block stillness bounds its mass from below, not above, so the cascade prediction depends on the prior’s 0.40 kg upper bound.The parameter belief also covers both domino masses, shared friction, wind force, and decay time; Table 2 reports posteriors after the two probes.Source: §C.4 · p. 25, Table 2 · p. 26
  • Two probes distinguish the blocks: the green block stays upright, while the gray block topples and moves approximately 12.1 cm into the original target region.The agent places each block approximately 25 cm downwind and holds the button for 2 s.Source: §C.4 · p. 24, Figure 7 · p. 25
  • The gust topples the gray block into the green one, which slides 9.9 cm against a predicted 10.1 ± 2.3 cm and rests flat in the relocated patch.The agent places the green block at 37.5 cm and the gray block upwind at 20 cm before activating the fan.Source: Figure 9 · p. 25, §C.4 · p. 25, Figure 8 · p. 25

C.5 DISCUSSION

  • The cascade is one run on one bench, and executing the loop there depends on perception, calibrated contact skills, and block recovery.The model combines fitted wind with existing contact dynamics to support a cascade in a new target layout.Source: §C.5 · p. 25

D TASK GOALS AND AGENT PROMPT

This appendix documents the EMPIRIC agent’s task prompts and system guidance, covering residual-model implementation, observation-driven state, and model testing. It also describes predicate design, latent-state use, execution conventions, and validation cautions.

D.1 TASK GOALS

  • Domino tasks require the blue dominoes to trigger a green-to-purple cascade with as few as possible, while Bridge tasks require a stable n-shaped structure on marked sites.The Domino test layout turns at a right angle, unlike the training line; the Bridge test span uses four blocks rather than three.Source: §D.1 · p. 27
  • Balloons tasks require freeing balloons to hold a box motionless within a specified height band, while Boil tasks require boiling water without spills and turning off the burner afterward.The Balloon test changes the box’s target band and adds a fourth clip; the Boil test uses two jugs.Source: §D.1 · p. 27
  • Fan tasks require blowing a ball to a target while keeping it on visible platforms, then leaving all fans off as it settles within the specified position tolerance.The test task changes the target position from the training task.Source: §D.1 · p. 27

D.2 FIRST MESSAGE OF A TASK

  • The Bridge opening message gives the goal alongside the ledger, observation, model status, journal, attempts record, and vocabulary; initially, no model is available and the visible-physics simulator is instructed to be fitted before test actions.The message reports that simulation uses real skill controllers with hidden mechanisms disabled, so reach, grasp, and collision checks already work.Source: §D.2 · pp. 27–28

D.3 SYSTEM PROMPT

  • The prompt prioritizes minimizing environment steps and resets, recovering in place when possible, and treating environment-certified success rules as authoritative.Train levels allow evidence collection, while test-level skill calls are refused until a simulator loads and declares residual features.Source: §D.3 · pp. 29–30
  • The prompt distinguishes supplied environment predicates from model-memory beliefs and warns that noisy-frame predicates can disagree with the evaluator’s true-state judgment.Non-robot object positions have Gaussian noise with sigma 0.005 m and orientations with sigma 0.02 rad; discrete features, switch states, and robot state are exact.Source: §D.3 · p. 30
  • The system prompt directs the agent to use recordings and sandbox computation, fit and validate a model against training evidence, rehearse plans, then act and update from observed divergences.Before actions that could finish or lose a level, it calls for replaying the full plan under multiple trials and with contact checking.Source: §D.3 · p. 30
  • The prompt says to investigate rejected fits by checking trajectory errors and residuals, comparing alternative dynamics structures and parameters, and probing only when model disagreement could change the next action.Probes should be low-cost, distinguishable from observation noise, and preserve future choices where possible.Source: §D.3 · p. 30
  • A plan uses one typed skill invocation per line; expectations are checked after execution, and a mismatch is reported as a divergence that normally stops the remaining plan.Skill parameters must be exact, and the optional integer to Wait counts environment steps, not seconds.Source: §D.3 · p. 31
  • `sim.fit()` estimates declared parameters from recordings, `sim.validate()` replays them at deployed values, and separate fit and validation calls compare candidate hypotheses on identical data.Validation includes recordings rejected by a robust fit; residual reports help locate errors.Source: §D.3 · p. 32
  • Refined plans can be rehearsed and tested under physical-parameter sweeps or belief draws, but these checks are conditional on the model; only a live-environment `WIN` certifies completion.For replay verdicts, provide action labels when the evaluator replays actions and inspect its note to determine what was scored.Source: §D.3 · p. 32
  • Implement missing dynamics in a `BaseSimulator` subclass exported as `RESIDUAL_ENV`, using `_domain_specific_step`; each primitive action advances base physics before calling that hook.Forces applied by the hook take effect on the following physics step; use engine constraints for coupled bodies and preserve added state across resets and branches.Source: §D.3 · pp. 32–33
  • Use `update_model_state` for pure, observation-driven memory updates once per primitive action; runtime state is an inferred estimate, not measured truth or a particle filter.Memory persists across prediction, resets, and planning branches, and the first observation initializes it without advancing it.Source: §D.3 · pp. 33–34
  • Declare learnable parameters with finite plausible bounds, and retain raw recorded features for fitting because the fit accounts for observation noise and the model noise floor.Parameters must affect scored features to be identifiable; a rejected fit alone does not distinguish structural, parameter, state, or fitting causes.Source: §D.3 · p. 34
  • Export `LEARNED_PREDICATES` as classifiers that receive a state and bound objects and return a boolean.For geometric conditions, transform learned local offsets by object orientation and choose thresholds that separate recorded positive and negative examples.Source: §D.3 · pp. 33–34
  • Define predicates for outcomes used by skill expectations, divergence checks, and `Wait` targets.The prompt advises defining predicates for the outcomes the agent relies on.Source: §D.3 · p. 34
  • Share each physical threshold with its mechanism and keep completion thresholds reachable within the model's output range.Source: §D.3 · p. 34
  • After predicate edits, call `sim.predicates()` to load definitions and inspect whether each grounding holds, changes, or latches in recordings.Source: §D.3 · p. 34
  • Supplied environment predicates and invented predicates remain distinct even when they share a name.Source: §D.3 · p. 34
  • A classifier can accept a keyword argument named exactly `latent` to read inferred model memory, which `sim.predicates()` reconstructs over recordings before scoring.Source: §D.3 · p. 34
  • Treat latent-dependent classifier outputs as model-dependent, and prefer observable classifiers when their readings already carry the needed signal.Source: §D.3 · p. 34
  • The agent runs in a local sandbox with listed tools, and all file operations are restricted to the current workspace.The prompt directs the agent to consult workspace documentation and reference files for system APIs before writing code.Source: §D.3 · p. 34

Related papers

Loading 2609.35047v1…