PaperlayerView exact revision on arXiv

Source-linked AI summary

ManiPhysicsBench: Physics-Based Assessment of Object Preservation in VLA Manipulation

Sangwu Park, Yeonjun In, Wonjoong Kim, Sungwon Kim, Sein Kim, Chanyoung Park

arXiv:2610.02802v1cs.RO

TL;DR

Rigid-body manipulation benchmarks score task completion without establishing whether objects are preserved. ManiPhysicsZoo and solver-based assessment support a benchmark that checks grasp forces against object-specific damage thresholds. Public VLA checkpoints show a task-success/safe-success gap, while continuous object-specific supervision raises safe success but lowers task success and generalizes only to a limited extent.

  • Problem

    Task success in rigid-body benchmarks does not establish whether a VLA preserves an object while completing manipulation.

    Source: §1 · p. 1

  • Method

    The paper combines reusable, literature-supported object assets with solver-based assessment of recorded grasp forces against grasp-specific damage thresholds.

    Source: §1 · p. 2

  • Results

    Public VLA checkpoints show a substantial gap between task success and safe success, while object-specific continuous gripper labels yield higher safe success but lower task success and limited generalization.

    Source: §1 · p. 2, §7 · p. 9

  • Takeaways & Limitations

    Object-specific continuous supervision produces more object-dependent gripping, but does not ensure reliable preservation across distinctions not supplied during training.

    Source: §1 · p. 2, §6.3 · p. 9

    Object-specific supervision alone does not ensure reliable object-preserving manipulation; paper and portion cups still have no safe successes despite successful task completions.

    Source: §6.3 · p. 9

Abstract

from arXiv · show

Vision-language-action (VLA) models aim to perform diverse manipulation tasks, but task success in existing rigid-body benchmarks does not indicate whether they preserve objects. We introduce ManiPhysicsZoo, which consolidates literature-supported material properties, 3D meshes, and supporting references into reusable object assets. Using these assets, a solver-based assessment computes grasp-specific damage thresholds from object geometry, material properties, and recorded grasp conditions and compares them with recorded contact forces to assess potential deformation and fracture. Building on these components, ManiPhysicsBench evaluates object preservation in LIBERO and SimplerEnv across three physics axes and three difficulty levels. Public VLA checkpoints show a substantial gap between task success and safe success, defined as task completion while preserving the object. Their gripper commands concentrate near full opening and closure, with largely similar aggregate distributions across objects, consistent with binary gripper supervision. We examine how object-specific continuous gripper labels change model behavior by retraining a VLA model. The retrained model shows more object-dependent gripping and higher safe success, but lower task success and limited generalization of object-preserving behavior.

1 INTRODUCTION

Task completion in rigid-body benchmarks does not show whether a policy preserves objects. The paper introduces reusable physical assets and solver-based damage assessment, then evaluates public VLAs and object-specific gripper supervision.

1 INTRODUCTION

  • The paper asks whether VLAs can account for object properties while completing manipulation tasks without damage.Real-world repeated trials incur replacement costs for fractured or permanently deformed objects, and damaged objects may pose safety risks.Source: §1 · p. 1
  • Existing rigid-body benchmarks do not model damage, and available resources for object-specific preservation criteria are not unified or verifiable.The assessment must also translate rollout forces into predictions of deformation or fracture.Source: §1 · p. 2
  • MANIPHYSICSZOO consolidates object-specific properties, meshes, and references, while solver-based assessment compares grasp-specific damage thresholds with recorded forces.The benchmark uses these components to evaluate preservation in LIBERO and SimplerEnv across three physics axes and three difficulty levels.Source: §1 · p. 2
  • Public VLA checkpoints show a substantial gap between task success and safe success; continuous object-specific labels yield higher safe success but lower task success and limited generalization.Their gripper commands cluster near fully open and fully closed states, with largely similar aggregate distributions across objects.Source: §1 · pp. 2–3

2 RELATED WORK

Related work evaluates manipulation, physical understanding, safety, or force-aware behavior, but leaves gaps in a common, physically grounded assessment of object preservation during task execution.

2 RELATED WORK

  • Manipulation benchmarks model objects as rigid bodies, so task success alone does not indicate whether an object was preserved.Table 1 frames comparison across simulation, preservation, physical grounding, and sourced object libraries.Source: §2 · p. 3, Table 1 · p. 3
  • Physical-understanding benchmarks use question answering rather than executed manipulation, while deformable environments target shape change without preservation criteria.DefGraspSim models grasp-induced stress and deformation, but uses fixed material parameters and does not include an arm performing a manipulation task.Source: §2 · p. 3
  • Robot safety benchmarks share the distinction between task and safe completion, but do not provide a common physically grounded object-preservation verdict.SafeVLA-Bench focuses on scenario safety, while OopsieVerse uses manually annotated damage criteria.Source: §2 · p. 3
  • Force- and tactile-aware methods adapt contact behavior and demonstrate it on real robots, but use dedicated protocols and visual post-trial inspection rather than a common criterion.Their real-hardware evaluations are subject to cost and safety constraints.Source: §2 · pp. 3–4
  • MANIPHYSICSBENCH combines reusable assets, grasp-specific solver-based damage assessment, and evaluation of public VLA checkpoints on existing manipulation platforms.Its judgments compare thresholds derived from object properties and recorded grasp conditions with rollout contact forces.Source: §2 · p. 4

3 PRELIMINARIES

The evaluation represents each policy episode as a rollout and extracts sustained parallel-jaw contacts as grasps, using recorded contact positions, directions, and forces.

3 PRELIMINARIES

  • A task instance combines an object, scene, language instruction, and platform; policy actions in the platform’s rigid-body simulator produce a rollout.The framework extracts a variable number of grasps from each rollout.Source: §3 · p. 4
  • A grasp is an interval in which both parallel-jaw gripper jaws contact the object for at least 0.2 s.Each grasp records contact position, compression direction, and the largest force sustained over the interval.Source: §3 · p. 4
  • Position-servo control turns gripper opening and closure targets into contact force, and the assessment assumes standardized silicone pads.The model accounts for force reduction as the pads and object deform.Source: §3 · p. 4

4 PHYSICALLY GROUNDED PRESERVATION ASSESSMENT

The framework combines a traceable object library with post-rollout solver-based assessment of grasp-specific damage, then applies both to benchmark tasks.

4.1 MANIPHYSICSZOO

  • MANIPHYSICSZOO selects objects whose size and graspable region suit common manipulators, whose gripping can cause irreversible damage, and whose appearance provides relevant cues.Its assets cover 91 objects; towels are excluded because ordinary gripping deformation does not compromise their intended use.Source: §4.1 · pp. 4–5
  • The library covers brittle fracture, soft-object deformation, and thin-walled container deformation, with meshes checked and adjusted for mechanical modeling.Missing structural features such as hollow-object wall thickness are added, and dimensions are scaled using published specifications or measurements.Source: §4.1 · p. 5
  • Material parameters come from reported measurements, datasheets, standards, or calibration to published tests, with each value’s test context recorded.The egg-shell model predicts measured half-eggshell breaking loads within 8–21% in a separate test without refitting.Source: §4.1 · p. 5
  • Each object asset stores its name, mesh and physical structure, material properties with measurement context and validation status, and supporting references.This representation is designed to support traceability and reuse across policies, robots, and simulation environments.Source: §4.1 · p. 6

4.2 SOLVER-BASED ASSESSMENT

  • Because rollout simulators do not model deformation or fracture, the assessment records contact positions, directions, and forces and associates each extracted grasp with its mesh region.It assesses potential damage after the rollout using the corresponding object asset.Source: §4.2 · p. 6
  • For each grasp, a solver uses object geometry, material properties, and contact configuration to find the smallest force reaching the object-specific damage criterion.The recorded force is compared with the threshold, adjusted for position-servo load reduction; a grasp is flagged as potentially damaging when its force exceeds that threshold.Source: §4.2 · p. 6

5 MANIPHYSICSBENCH

ManiPhysicsBench combines reusable, literature-supported object assets with solver-based damage assessment to evaluate preservation across LIBERO and SimplerEnv. Tasks are grouped by physical constraint and difficulty, while episode-level metrics separate task completion from preservation.

5.1 BENCHMARK CONSTRUCTION

  • The benchmark uses ManiPhysicsZoo assets and solver-based assessment to evaluate object preservation in LIBERO and SimplerEnv, including on rigid-body simulators.Tasks are organized across three physics axes and high, medium, and low difficulty levels.Source: §5.1 · p. 6
  • Difficulty levels reflect how narrow the gripping-force range is for lifting an object without damage, while size, grasp geometry, goals, and motion are kept as similar as possible.Low-difficulty objects are rarely damaged and serve as high-strength controls; Table 2 notes that axis C uses different paper cups across platforms because gripper opening ranges differ.Source: §5.1 · p. 6, Table 2 · p. 6

5.2 EVALUATION PROTOCOL AND METRICS

  • Table 3 reports SR, Safe SR, and PR; its Average covers High and Mid objects because Low objects are rarely damaged.Overall aggregates public checkpoints, and Relabel denotes object-specific gripper supervision.Source: Table 3 · p. 7
  • An episode counts as preserved only if none of its assessed grasps damages the object, so one damaging grasp is sufficient to fail preservation.Task completion follows the original task's success criterion.Source: §5.2 · p. 7
  • SR measures task completion, Safe SR measures completion while preserving the object, and PR measures preservation among completed tasks.PR is undefined when no episode succeeds; metrics are reported by physics axis and difficulty level.Source: §5.2 · p. 7

6 EXPERIMENTS

Experiments on SimplerEnv and LIBERO examine safe task completion, gripper behavior, and object-specific continuous supervision. Public checkpoints often complete tasks without preserving objects, while retraining shifts gripping behavior but does not remove the trade-off or generalization limits.

6.1 MAIN RESULTS

  • On SimplerEnv, public checkpoints average 48.3% SR but 13.5% Safe SR across the assessed objects, with PR of 27.9%.For the egg and potato, nine of ten checkpoints have 0.0% Safe SR despite SR reaching 91.7%.Source: §6.1 · p. 8
  • Mean PR is 15.1% on LIBERO versus 27.9% on SimplerEnv, although the portion cup reverses the ordering at 57.1% versus 7.1%.For the portion cup, LIBERO π0 preserves 85.7% of successful episodes; its damage threshold depends strongly on grasp location.Source: §6.1 · p. 8, Table 4 · p. 8

6.2 GRIPPER ANALYSIS

  • Gripper commands concentrate near fully open and fully closed states, with similar aggregate distributions across axis A objects despite different damage thresholds.The pattern resembles binary open/close supervision, which provides no explicit object-dependent closing targets.Source: §6.2 · p. 8, Figure 2 · p. 8

6.3 OBJECT-SPECIFIC GRIPPER SUPERVISION

  • The retraining procedure replaces selected binary gripper labels with continuous closing targets derived from object-specific physical constraints and force estimates.QwenGR00T is retrained on relabeled Bridge data; the run differs from the public checkpoint in data mixture and configuration, so it does not isolate label replacement.Source: §6.3 · pp. 8–9
  • Object-specific supervision yields more differentiated gripping and higher preservation, but lower task completion across benchmark objects.On the low-friction ice-cube task, the models achieve similar task success and retain near-full closure during contact, consistent with object-dependent rather than uniformly weaker gripping.Source: §6.3 · p. 9, Figure 3 · p. 9, Table 5 · p. 9
  • Paper and portion cups in axis C still have no safe successes despite task completions, and generic cup annotations do not distinguish paper from plastic cups.The results indicate limited generalization beyond distinctions explicitly supplied in supervision; object-specific supervision alone does not ensure reliable preservation.Source: §6.3 · p. 9

7 CONCLUSION

ManiPhysicsZoo and ManiPhysicsBench combine reusable object assets with solver-based assessment to measure object preservation during manipulation. Public VLA checkpoints show a gap between task success and safe success, while the paper documents the benchmark’s scope, resources, and limitations.

7 CONCLUSION

  • Public VLA checkpoints show a substantial gap between task success and safe success; retraining yields more object-dependent gripping and higher safe success, but lower task success and limited generalization.Their gripper commands concentrate near fully open and fully closed states, with largely similar aggregate distributions across objects.Source: §7 · p. 9
  • The assessment uses literature-supported material properties and damage criteria, but its damage predictions still require validation through physical robot experiments.It evaluates bilateral gripping after each rollout and does not capture damage from drops or impacts or how deformation affects subsequent contacts and policy actions.Source: §7 · p. 9

AI USE STATEMENT

  • Generative AI assisted with implementation, mesh generation, language editing, and literature searches; authors reviewed the resulting code, assets, and collected information.An LLM also helped extract manipulated objects and annotate potential damage risk from Bridge instructions; extraction and classification errors may remain.Source: AI USE STATEMENT · p. 11

ETHICS STATEMENT

  • Policy evaluations and mechanical assessments use simulation, existing datasets, checkpoints, and object assets; simulation assessments are not a guarantee of safe physical deployment.The paper documents asset provenance and material-property sources, while third-party resources remain subject to their original licenses.Source: ETHICS STATEMENT · p. 11

REPRODUCIBILITY STATEMENT

  • The accompanying repository provides evaluation, solver, relabeling, training, and aggregation code and configurations, along with dependencies, instructions, seeds, solver inputs, and outputs.It also points to tables and appendices documenting checkpoints, benchmark assets, material-property justification, and procedures.Source: REPRODUCIBILITY STATEMENT · p. 11

A.1 OBJECT COLLECTION

  • ManiPhysicsZoo contains 91 object records spanning household, industrial, and laboratory objects, each combining a visual asset with geometry, material information, and supporting evidence.The collection includes 50 household, 33 industrial, and 8 laboratory objects.Source: §A.1 · p. 17

A.2 OBJECT ASSETS AND EVIDENCE SCHEMA

  • Object records organize names, meshes and geometry, material properties, and supporting references, with provenance and verification information documented in the asset schema.Supporting records distinguish measured properties from derived, transferred, calibrated, or assumed values and make missing inputs and pending verification explicit.Source: §A.2 · p. 17

A.3 LITERATURE-BASED JUSTIFICATION

The literature-based justification distinguishes evidence for material responses from validation of whole-object damage loads. It documents object-specific criteria and comparisons while noting that model uncertainty affects damage verdicts.

Evidence levels and parameter provenance

  • Evidence levels distinguish quantitative reproduction or comparison of published responses, complementary material-and-geometry or transfer-based support, and unresolved provenance or test-correspondence gaps.Calibration responses are distinguished from additional test conditions, while numerical convergence and literature-response agreement are assessed separately.Source: §A.3 · p. 22

Interpreting the benchmark evidence paths

  • Apart from low-difficulty, high-strength walnut and plum controls and material-dominated containers, benchmark objects have Level 1 quantitative support for their specified mechanical responses.The cited checks cover eggshell breaking loads, tomato compression reaction, potato stress at observed fracture strains, and reproduction of the strawberry bioyield calibration target.Source: §A.3 · p. 22
  • Walnut remains undamaged even against the lowest reported rupture load: its maximum recorded per-jaw force was 16.0 N, or 0.40 times that load.This supports walnut as a high-strength control despite legacy shell parameters lacking a traceable source; plum is also used as a low-difficulty control.Source: §A.3 · p. 22

Egg: axis A, High

  • The egg’s 5.05 N median benchmark damage reaction per jaw lies within the modeled 1.5–6.3 N inner-crack initiation range, rather than contradicting higher break loads.In 25 sampled grasp analyses, the criterion was first reached on the inner surface beneath the pad, while the outer surface remained at 14–22% of tensile strength.Source: §A.3 · p. 23
  • For half-eggshell compression, predicted outer-crack loads were 57.6, 55.8, and 37.8 N, within +8 to +21% of measured breaking forces and about 3% of source FE predictions.Shell properties and tensile strength were benchmark values obtained from separate tensile tests; none of these loads was used to set them.Source: §A.3 · p. 23, Table 8 · p. 23

Cherry tomato: axis A, Medium

  • 143.46–146.61 N modeled tomato reactions differed by +1.67–3.91% from the reported 141.10 N at an observed rupture displacement.This supports the axial reaction response, not rupture-onset prediction or direct validation of the 40 mm asset’s grasp threshold.Source: §A.3 · p. 23

Boiled potato: axis B, High

  • 54.37 and 54.10 kPa potato stresses differed by +8.30% and −8.45% from literature values at observed fracture strains, without refitting to those stresses.The benchmark treats tissue with accumulated equivalent plastic strain PEEQ ≥0.002 in at least 1% of initial volume as unacceptable modeled permanent deformation.Source: §A.3 · p. 24

Strawberry: axis B, Medium

  • The strawberry coupon reproduces its literature-based calibration target, but this does not independently establish whole-fruit damage error or the complete post-yield response.The benchmark uses the resulting effective cortex model with the same 1% damaged-volume allowance as the potato.Source: §A.3 · p. 24

B EVALUATION SETUP

The evaluation uses public VLA checkpoints compatible with each platform and reports results separately for the WidowX and Panda robots. Each model-object pair receives 12 rollouts under the platform’s original task protocol.

Platform-specific evaluation

  • WidowX and Panda results are reported per platform because their gripper hardware and control differ, and platform differences are not treated as a model ranking.Evaluated public checkpoints are those compatible with each platform as released.Source: B EVALUATION SETUP · p. 26

SimplerEnv checkpoints

  • The SimplerEnv evaluation includes nine public VLA checkpoints: CogACT, GR00T N1.6 and N1.7, InternVLA-M1, LangForce, Octo, π0, QwenGR00T, SpatialVLA, and VLA-JEPA.Each model and object receives 12 rollouts under the original task protocol.Source: B EVALUATION SETUP · p. 26

LIBERO checkpoints

  • The LIBERO evaluation includes 12 public VLA checkpoints, including ABot-M0, DreamVLA, GR00T N1.7, MolmoAct2, Octo, OpenVLA-OFT, π0, and π0.5.The full list also includes RIPT-VLA, SmolVLA, UniVLA, and VLA-Adapter; each model and object receives 12 rollouts.Source: B EVALUATION SETUP · p. 26

B.1 TASK ADMISSION

Tasks are admitted only after an oracle demonstrates successful, object-preserving manipulation. The appendix specifies the evaluated baselines and the position-servo, force-conversion, and contact-pad assumptions used to assess their grasps.

B.1 TASK ADMISSION

  • An oracle varies grip force with prescribed arm motions, admitting a task only if it completes and passes solver-based preservation assessment.Objects are selected for each platform’s gripper range, and feasibility is checked separately for Panda in LIBERO and WidowX in SimplerEnv.Source: §B.1 · p. 26

B.2 BASELINE MODELS

  • The baseline set includes policies using diffusion or flow-based action generation, continuous action sequences, and diverse vision-language or world-model architectures.Examples include CogACT, GR00T, π0, ABot-M0, and DreamVLA.Source: §B.2 · p. 26

C GRIPPER CONTROL

  • Position-servo control matches current VLA opening and closure commands; the simulator’s resulting contact forces are recorded for damage assessment.Because physical gripper responses and training-set grip forces are not assumed identical to the simulation, the simulator response is used as a common reference that can be calibrated toward hardware.Source: C GRIPPER CONTROL · p. 29

C.1 CONTROLLER SETTINGS

  • Host controller settings are inherited, and an obstructed jaw target creates drive force through the position controller.The settings are shared across policies within each platform; force limits apply per jaw actuator, with damping implemented differently on WidowX and Panda.Source: §C.1 · pp. 29–30, Table 10 · p. 30

C.2 NOMINAL FORCE–CLOSURE RELATION

  • The quasi-static, unsaturated approximation maps per-jaw closing position error to nominal drive force and supports force-target conversion to position offsets.At kp = 1000 N/m, a 1 mm per-jaw error corresponds to a nominal 1 N drive force; the relation also accounts for reduced servo load during deformation.Source: §C.2 · p. 30

C.3 PAD–OBJECT CONTACT MODEL

  • The mechanical assessment assumes compliant silicone pads, whose geometry and material properties are modeled alongside the recorded grasp and object geometry.Contact stress can vary with grasp location, surface curvature, and the compliance of the contacting bodies.Source: §C.3 · p. 30

C.4 SENSITIVITY TO THE POSITION GAIN

Rescaling recorded forces while holding trajectories and FEM damage loads fixed changes safe success, but policies’ commands do not track the resulting safe closure intervals. The solver appendix describes how grasp-specific damage thresholds are computed and where the conversion has limits.

C.4 SENSITIVITY TO THE POSITION GAIN

  • Halving kp raises Safe SR from 13.5% to 24.9% on SimplerEnv and from 1.9% to 3.4% on LIBERO, while doubling kp lowers it to 3.6% and 0.9%.Task success rates are 48.3% and 12.3%, respectively; Safe SR at 0.5kp is an upper bound because saturated forces are underestimated.Source: §C.4 · p. 30, Table 11 · p. 31
  • Median force-to-threshold ratios are 2.0 for the egg and 2.6 for the potato on SimplerEnv, 3.8 and 4.5 on LIBERO, and 6.0 for paper cups on both platforms.The strawberry on SimplerEnv and cherry tomato on LIBERO lie close to threshold; Figure 6 plots each successful episode’s largest ratio, with values above one indicating damage.Source: §C.4 · p. 31, Figure 6 · p. 31
  • Safe closure intervals exist for every analyzed grasp at all three gains, but policies remain near full closure rather than adapting commands to objects.At 0.5kp, some grasps become safe because full closure no longer exceeds the threshold, not because commands adapt to the object.Source: §C.4 · p. 31, Figure 7 · p. 32

D.1 OBJECT MECHANICS AND PHYSICS SOLVERS

  • Solver routes match damage modes: brittle objects use implicit FEM, soft objects use nonlinear constitutive models, and thin-walled containers use explicit dynamics.The modeled criteria include fracture-strength exceedance, irreversible deformation beyond an application-level volume tolerance, and permanent indentation or buckling.Source: §D.1 · p. 33

D.2 GEOMETRY AND MESHING

  • The contact meshes refine around recorded contact regions, with separate element choices for thin structures, volumetric objects, and skin–core models.Standard edge targets are 2 mm near contact and 4 mm away; the 2 oz paper cup uses a 3 mm object refinement target.Source: §D.2 · p. 33

D.3 CONTACT AND LOADING CONDITIONS

  • Each grasp is solved as a coupled object-and-pad contact problem under progressive compression, with evolving contact regions and traction distributions.Weak centering springs regularize initially free rigid-body modes and do not represent the robot’s position-servo gain.Source: §D.3 · p. 33

D.4 NUMERICAL SOLVERS

  • The numerical configurations distinguish solver routes, while loading ramps and increments are adjusted to meet checks over the interval supporting each damage decision.Load durations and numerical increments are solver settings, not damage tolerances.Source: §D.4 · p. 33, Table 12 · p. 34

D.5 DAMAGE THRESHOLDS AND SERVO COUPLING

  • Thresholds convert FEM criterion-crossing loads to recorded-force equivalents; nonmonotonic cases require a reachable stable branch, while nonconvergent cases use nearby same-object grasps.The fallback uses the median directly computed threshold from the five nearest grasps by contact-point location.Source: §D.5 · p. 34

D.6 COMPUTATION AND GPU ACCELERATION

The assessment is performed offline on recorded grasps, with reusable solver results and parallel execution reducing elapsed time. The accompanying relabeling pipeline selects object-specific episodes, derives force targets, and converts them to closing offsets.

D.6 COMPUTATION AND GPU ACCELERATION

  • Offline assessment reuses cached curves for matching geometry, material, contact, and solver settings; GPU acceleration targets sparse linear-system solves, while OpenRadioss runs on CPUs.When saved histories cover the loading path, servo coupling can be evaluated by postprocessing without another FEM solve.Source: §D.6 · pp. 34–35
  • 273 s for boiled potato and 531 s for strawberry are the median FEBio wall times per attempted grasp; a representative three-thread OpenRadioss cup solve takes 49.48 min.These are recorded configurations, not a controlled solver comparison; the reported statistics include unsuccessful FEBio attempts.Source: §D.6 · p. 35

E.1 INSTRUCTION-BASED EPISODE SELECTION

  • 8,836 episodes (16.61% of the dataset) are retained for target filtering, while 44,356 episodes keep their original labels.The candidate set contains 8,989 episodes across 74 force-target categories; the annotation workflow standardizes object names and filters for manipulation phrase and grasp interaction.Source: §E.1 · pp. 36–37, Table 13 · p. 37, Table 14 · p. 37

E.2 OBJECT-CONDITIONED FORCE TARGETS

  • When lifting and damage-load estimates are both available and the lifting requirement is lower, the target force is their midpoint; otherwise, size and mass assumptions estimate lifting force and a grip margin is added.Targets use category-level load proxies and a shared force cap; representative-object masses and proxies do not guarantee lifting or preservation for every instance.Source: §E.2 · p. 37

E.3 FORCE-TO-ACTION CONVERSION

  • The position-servo relation converts each category’s nominal per-jaw force target into a closing offset, with the numerical value in newtons equaling the offset in millimeters.This is a nominal conversion, not measured contact-force calibration; evaluation assessment uses recorded contact loads.Source: §E.3 · p. 38
  • Only selected episodes’ closing labels are replaced; arm actions, opening commands, and unselected episodes remain unchanged.Relabeling is applied in the data loader without rewriting the source dataset.Source: §E.3 · p. 38

E.4 TRAINING CONFIGURATION

The paper specifies the retraining setup, aggregation rules for executed gripper commands, and model-wise distribution displays. It also defines difficulty using a reference index of the safe gripping-force range and limits the reported FEM summaries to assessed grasps.

E.4 TRAINING CONFIGURATION

  • The retrained QwenGR00T uses Bridge-only data, VLM LoRA, and a newly initialized action model; its public-checkpoint comparison does not isolate label replacement because training data and configuration differ.The run stops at 20,318 updates, but evaluation uses the 20,000-update checkpoint while retaining a 100,000-update learning-rate horizon.Source: §E.4 · p. 38

F.1 EXECUTED CLOSURE AND AGGREGATION

  • Gripper commands are reduced to one per policy control step and mapped to a shared closure coordinate, with histograms including unsuccessful episodes and non-contact intervals.Contact-conditioned means include steps with bilateral object contact and average across episodes with such contact, including task failures; episodes without bilateral contact are excluded.Source: §F.1 · p. 40

F.2 MODEL-WISE DISTRIBUTIONS ACROSS PHYSICS AXES

  • Figures 9 and 10 show model-specific distributions across nine objects, with 12 episodes averaged equally per model-object histogram; their markers describe near-binary command mass and contact-conditioned ordering, not significance or safety.The figures retain High/Mid/Low objects within each physics axis; an ordering marker is absent when an object lacks bilateral-contact observations.Source: §F.2 · p. 40, Figure 9 · p. 41, Figure 10 · p. 42

G DIFFICULTY LEVELS

  • Difficulty uses the safe-force range relative to lifting force: D = 0 means assessed thresholds are not reached within the reference range, while larger D indicates a narrower damage-free lifting range.The index is a physical reference measure, not task success or a platform-independent contact-force claim; Table 16’s medians and D values describe only deduplicated grasps with valid, observed criterion crossings.Source: G DIFFICULTY LEVELS · p. 43, Table 16 · p. 43

H LOW-FRICTION ICE CUBE EVALUATION

The ice-cube evaluation compares public and retrained QwenGR00T models on the same SimplerEnv task and initial configurations. Its friction setting is a literature-informed proxy for pad contact.

H LOW-FRICTION ICE CUBE EVALUATION

  • The evaluation uses the SimplerEnv put-on-towel task with an ice cube and compares public and retrained QwenGR00T over the same 12 initial configurations.Task success is reported using the platform’s success criterion.Source: H LOW-FRICTION ICE CUBE EVALUATION · p. 44, Figure 11 · p. 44
  • The object and jaw surfaces use static and dynamic friction of µ = 0.12, chosen as a literature-informed proxy for rubber–ice contact.The passage cites Tada et al. (2023) as the reported low-friction regime informing this setting.Source: H LOW-FRICTION ICE CUBE EVALUATION · p. 44
  • Mean closure is averaged over bilateral-contact steps within each episode, then across episodes with such contact.Source: H LOW-FRICTION ICE CUBE EVALUATION · p. 44
Loading 2610.02802v1…