SimEX: Simulation-Integrated Robotics AutoResearch

1 INTRODUCTION

• Direct robot-code generation can depend on costly, embodiment- and environment-specific primitives, while physical autoresearch consumes robot time and raises safety concerns.Without such primitives, models must guess quantities such as camera-calibration errors or reachable grasps; these guesses can lead to catastrophic failure.Source: §1 · pp. 1–2
• Simulation offers a code-accessible way to test hypotheses safely, but learned capabilities must be grounded in real trials because simulators can be inaccurate.Source: §1 · p. 2
• SimEX develops capabilities through open-ended simulated experiments, then uses a few physical trials to generate improvement hypotheses and screen candidate repairs in simulation.The authors describe this two-stage design as robust to an imperfect simulator: broad exploration limits overfitting, and physical grounding corrects simulator errors.Source: §1 · p. 3
• 26 of 30 physical evaluation trials succeed with SimEX across three task families, compared with 3 for the strongest baseline.The evaluation uses four coding-agent types and only 10 minutes of on-robot interaction per task.Source: §1 · p. 3

2 METHOD

• The method first develops a toolbox in simulation, then concentrates physical trials in a final phase that screens repairs between rollouts.Concentrating trials allows an operator to initialize and monitor the robot in one session; initial capabilities are intended to make rollouts safer and more efficient.Source: §2 · p. 3

2.1 PROBLEM SETUP

• The toolbox is a persistent pair of perception/control primitives and usage guidance; the task-specific program is discarded, so learned knowledge persists through toolbox updates.Optimization leaves the low-level robot API fixed; perception uses raw RGBD images without privileged scene state or predefined task-level skills.Source: §2.1 · pp. 3–4

2.2 PREPARATION: CONSTRUCTING THE SIMULATION SANDBOX

• Preparation reconstructs an executable approximation of the workspace from an image and initializes the simulator, perception module, and robot-API guidance.The target is a usable starting point for optimization, not a precisely calibrated or photorealistic digital twin.Source: §2.2 · pp. 4–5

2.3 STAGE 1: OPEN-ENDED CAPABILITY DISCOVERY IN SIMULATION

• Stage 1 probes self-generated tasks and environment variations, using executable success predicates to find toolbox weaknesses without a fixed global metric.The agent chooses a systematic weakness and builds a focused training set, providing a stable objective for comparing toolbox revisions.Source: §2.3 · p. 5
• Stage 1 repeats probe–focus–optimize cycles, using capability-bank replay to guard against regressions and independent populations to diversify the search.Population members can share experiment descriptions and hypotheses, but not implementations; each maintains its own toolbox, task sets, and capability bank.Source: §2.3 · pp. 5–6

2.4 STAGE 2: SIMULATION-ASSISTED FEW-TRIAL ADAPTATION ON THE ROBOT

• Stage 2 turns each scarce robot rollout into simulated tests of distinct candidate repairs rather than immediately trying the first plausible edit on the robot.The agent aligns the simulator through fixed-policy replay, diagnoses the rollout, and forms testable repair hypotheses.Source: §2.4 · p. 6
• Repair screening tests candidates on the fixed executed script and simulated tasks, but the agent—not an automatic ranking rule—judges the evidence and selects a repair or keeps the current toolbox.The simulated tasks check whether a local fix generalizes beyond the observed failure, and the selected toolbox is used in the next robot rollout.Source: §2.4 · p. 6

3 EXPERIMENTS

• The evaluation combines controlled sim-to-sim comparisons, physical-robot tests, stage ablations, and comparisons across four coding agents.Sim-to-sim evaluation uses distinct sandbox and evaluation simulators, with the coding agent unable to inspect the evaluation simulator's state or source code.Source: §3 · p. 7, §3.2 · p. 8

3.1 EXPERIMENTAL SETUP

• The evaluation covers plate-to-tote, towel-folding, and barcode-scanning tasks on a dual-arm YAM robot, with seven held-out evaluation tasks per family.The task families span contact-rich manipulation, deformable-object control, and long-horizon object handling.Source: §3 · p. 6, §3.1 · p. 7

3.2 CONTROLLED SIM-TO-SIM EVALUATION

• 85% average success on barcode scanning, while no external baseline exceeds 18%; across all 21 tasks, SimEX matches or exceeds every external baseline.SimEX also achieves the highest average success in all three task families and outperforms both stage-level ablations in every family.Source: Figure 3 · p. 7, §3.2 · p. 8
• In barcode online adaptation, both methods sharing a Stage 1 toolbox begin near zero; one deployment rollout and repair sharply improve both, while sandbox guidance more than doubles the gain.The sandbox-guided advantage persists through the fifth rollout.Source: §3.2 · p. 8, Figure 4 · p. 8

3.3 REAL-ROBOT EVALUATION

• The physical evaluation gives each adaptive method five on-robot adaptation trials, then evaluates each method in 10 trials per task.The five adaptation trials amount to about 10 minutes of robot interaction per task; scenes are reset by a human operator between trials.Source: §3.3 · p. 8
• SimEX succeeds in 26 of 30 physical evaluation trials, while no baseline succeeds in more than 3; its advantage spans all three task families.Every baseline fails all trials on two-object barcode scanning; the evaluated tasks are plate to tote, towel folding, and barcode scanning.Source: §3.3 · p. 8, Table 1 · p. 8

3.4 ABLATIONS AND FURTHER ANALYSIS

• Every fine-grained ablation reduces performance, with the largest drops from removing Stage 1 population diversity and capability-bank replay.The ablated Stage 2 mechanisms are repair alternatives and hypothesis generation.Source: Figure 5 · p. 8, §3.4 · p. 9
• Across coding agents, SimEX remains strongest and leads the best external baseline by more than 50 percentage points; replacing GPT-5.5 with GPT-6 Astra raises SimEX from 67% to 82%.The comparison repeats the complete seven-task barcode-scanning evaluation with Opus 5, GPT-5.5, and GPT-6 Astra in place of Fable 5.1.Source: §3.4 · p. 9, Figure 6 · p. 9

4 RELATED WORK

• The paper situates SimEX among four areas: LLM-based robot control, agentic optimization, generative simulation and environment design, and sim-to-real transfer.The passage cites representative work in each area.Source: §4 · p. 9
• Unlike those lines of work, SimEX uses simulation for a coding agent to build a reusable code toolbox and to diagnose failures and screen repairs between physical trials.Source: §4 · p. 9
• The paper discusses each of the four related-work lines in Appendix E.Source: §4 · p. 9

• Across controlled sim-to-sim benchmarks and physical-robot tasks, SimEX outperforms direct generation, deployment-only iteration, simulation-only transfer, and simpler simulated search.Source: §5 · p. 9
• Simulation serves as a laboratory for developing and repairing capabilities, with transfer occurring through knowledge and procedures rather than a policy.Source: §5 · p. 10
• The evaluation uses one robot platform, and a human still resets the scene between Stage 2 physical trials.Source: §5 · p. 10

A.1 HYPERPARAMETERS

• Default settings cover population size, per-process pretraining, deployment-trial budget, and evaluation protocol, with corresponding phases matched across methods.Source: Table 2 · p. 14, §A.1 · p. 14

A.2 POLICY-WRITING PROMPT

• Every method uses a shared one-shot, text-only code-as-policy prompt containing the current skill guide, toolbox source, task description, and fixed output format.Source: §A.2 · pp. 14–15

A.3 AGENT THINKING TIME AND TOKEN USAGE

• For barcode-task toolbox optimization, reported agent costs include thinking time and output tokens; optimized-toolbox inference takes around 30 seconds per task.Source: §A.3 · p. 15, Table 3 · p. 15

B EVALUATION TASK DETAILS

• The three task families use the same dual-arm YAM platform and workspace enclosure, while task-specific objects and fixtures create different manipulation challenges.Source: Figure 7 · p. 15, Figure 8 · p. 15, Figure 9 · p. 16

B.1 PLATE TO TOTE

• Plate-to-tote tasks test geometry-aware, collision-free grasping and transport, including flat-plate pickup, multiple reaches, and shifted plate or tote layouts.Each delivered plate must be lifted at least 5 cm and placed inside the tote; unnamed plates must stay outside.Source: §B.1 · pp. 15–16

B.2 TOWEL FOLDING

• Towel tasks vary initial pose and require shifting, shrinking, or folding a deformable towel despite sensitivity to grasp location, drag path, and accumulated error.The rotated half-fold also requires a flat, half-fold shape; piled configurations fail.Source: §B.2 · p. 16

B.3 BARCODE SCANNING

• Barcode tasks require object discovery, feasible grasping, label presentation to a fixed reader, and subsequent stowing, including hidden labels and multi-object sequences.Tote walls, a narrow scanner window, and clutter constrain approaches and manipulation order.Source: §B.3 · p. 17

C BASELINE DETAILS

• Methods share the low-level API, initial toolbox, observations, agent settings, task seeds, evaluation protocol, and initial sandbox where applicable, isolating how they use simulation and robot experience.Source: C BASELINE DETAILS · p. 17
• An asterisk marks a baseline reimplemented to preserve its central mechanism under the shared interface and budget, not extra information or a relaxed evaluation.Source: C BASELINE DETAILS · p. 17
• Baselines range from direct code generation and deployment-only edits to simulation-only transfer, simulated skill-library repair, fixed-target search, and direct rollout repair.SimEX direct repair omits simulator replay, hypothesis generation, and candidate screening.Source: C BASELINE DETAILS · p. 18

• Each table cell reports the number of successful evaluation trials out of 25 scheduled trials.Source: §D.1 · p. 18, Table 4 · p. 18
• Boldface identifies the highest success proportion in each row.Source: §D.1 · p. 18
• The table presents controlled sim-to-sim results broken down by task.Source: Table 4 · p. 18

D.2 PER-TASK ABLATION RESULTS

• Stage 1 ablations use one optimization process instead of four or remove capability-bank replay; Stage 2 ablations remove repair alternatives or hypothesis generation.Source: §D.2 · p. 18
• Each ablation setting uses five runs per task and five evaluation trials per run, with per-task counts reported in Table 5.Source: §D.2 · p. 18, Table 5 · p. 19

D.3 PER-TASK RESULTS ACROSS CODING AGENTS

• Table 6 reports barcode-scanning results for the all-Fable 5.1 main-experiment configuration, while Table 7 reports the end-to-end Opus 5 configuration.Source: Table 6 · p. 19, Table 7 · p. 19
• Tables 6–9 provide per-task results for seven tasks, with five independent runs and five evaluation trials per run for each configuration.Source: §D.3 · p. 19
• Across coding agents, SimEX consistently outperforms the strongest baseline methods by a large margin.Source: Figure 11 · p. 19

• For SimEX, the agent writes one task program before execution; the online policy repeatedly emits end-effector actions from camera-derived object locations and proprioception.Source: §D.4 · p. 20
• SimEX makes one model call before execution, taking about 30 seconds, and all three generated programs complete validation rollouts.Execution itself contains no model calls.Source: §D.4 · p. 20
• None of the online policy’s 15 rollouts completes the task within the allotted wall-clock budget.Source: §D.4 · p. 20
• The online policy makes about 30 model calls per rollout and waits 40 minutes for responses, over 50 times longer than its action-execution time.Source: §D.4 · p. 20

• Prior LLM-based robot-control methods generally assume that capabilities, feedback channels, or optimization interfaces are already specified; SimEX studies constructing and calibrating them through experimentation.Source: E EXTENDED RELATED WORK · p. 20
• Unlike prior embodied autoresearch examples, SimEX develops a robot capability library through open-ended simulated autoresearch and calibrates it with deployment feedback.Its simulated stage assumes no fixed task distribution, prescribed environment variations, or ground-truth evaluation metric.Source: E EXTENDED RELATED WORK · p. 21
• Generative simulation and environment-design methods mainly use generated content to train policies, whereas SimEX generates target-workspace task variations and success predicates to stress-test and expand a persistent symbolic code toolbox.Source: E EXTENDED RELATED WORK · p. 21
• Unlike simulation-pretraining methods followed by limited real-world adaptation, SimEX returns to simulation between physical trials to diagnose failures and screen code repairs.Source: E EXTENDED RELATED WORK · p. 21