Pith.

Artificial Intelligence · arXiv:2610.16782 v2 · submitted 12 Oct 2026

LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints

Ming Shin Cindy Sung, Duan Leijiahao, Haoxiang Wang, Boning Tan, Liangna Fu, Chao Tan, Yafu Li, Yulong Chen

show 4 more authors

Chaoqun Wan, Zhen Huang, Shuxian Liang, Xian-Sheng Hua

The paper says

“Industrial manipulation in flexible manufacturing, e.g., high-mix, low-volume (HMLV) electronics assembly, requires robot policies to adapt to changing products and production conditions.”

more of the abstract

Such settings impose industrial process constraints, including product variants, moving material flows, finite execution windows, and tight geometric tolerances. However, existing evaluation frameworks do not systematically capture these constraints when assessing generalist robot policies. To address this gap, we introduce LIBERO-Industry, a novel benchmark built on the widely used robot learning benchmark LIBERO, that translates the aforementioned industrial constraints into controllable task conditions and measurable success criteria. Specifically, LIBERO-Industry comprises three classic task scenarios: Static-to-Dynamic (S2D) loading, Dynamic-to-Static (D2S) unloading, and Dynamic-to-Dynamic (D2D) transfer. The suites contain 216 tasks in total that systematically vary object geometry, conveyor speed, interaction-window duration, and geometric acceptance region size. Extensive experiments across five generalist robot policies (including vision-language-action models and world-action models) show that performance progressively degrades as industrial constraints are introduced and combined. Under a shared fine-tuning and evaluation protocol, the LIBERO-Industry co-training condition achieved higher observed target-transfer completion than the original-LIBERO condition.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 14 Oct 2026 · of version 1 of the paper; version 2 is newer · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: major revision.

A well-designed industrial robot benchmark that fills a real gap, but the key sim-to-real result is confounded by data quantity, not just content. This line is wrong

2 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the references say.

The paper says, at §4.4, Appendix E.4

The review did not quote this passage.

Pith says: Objection 1 of 2 · major

The sim-to-real comparison (S1 vs. S2) confounds simulation data source with data quantity. S1 uses 1,724 original-LIBERO demonstrations (313,102 frames), while S2 uses 1,412 LIBERO-Industry demonstrations (620,660 frames). Both use a 9:1 simulation-to-real frame sampling ratio, so S2 injects approximately twice as many simulation frames per training step.

more of objection 1

The paper states the comparison 'isolates the simulation-data source and corpus size' (Appendix E.4), but corpus size in terms of frames is not held constant. The observed advantage of S2 (34.7% vs 26.4% PCR_T at 0.03 m/s; 23.6% vs 4.2% at 0.05 m/s) cannot be confidently attributed to the industrial content of LIBERO-Industry; it may simply reflect more simulation training data. The authors acknowledge this limitation (Appendix E.4 final sentence) but do not provide an additional controlled experiment or a quantitative analysis that bounds the effect. This weakens the central claim that co-training with LIBERO-Industry yields better transfer. The authors should either (a) repeat the comparison with matched total simulation frames (e.g., subsample LIBERO-Industry or upsample original LIBERO), or (b) clearly rephrase the claim as 'more simulation data, including industrially relevant data, leads to better transfer' and discuss the confound in the main text.

The paper says, at §4.1, Appendix C.3

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.78):

“Numeric evaluator seeds were not retained; accord- ingly, we report the observed proportions descriptively and do not claim a prespecified confidence-interval analysis.”

Pith says: Objection 2 of 2 · major

Reproducibility of the evaluation is compromised by the failure to retain numeric evaluator seeds. The paper states that for the cumulative diagnostic stress progression (D0–D6) 'numeric evaluator seeds were not retained; accordingly, we report the observed proportions descriptively and do not claim a prespecified confidence-interval analysis' (Appendix C.3).

more of objection 2

While the main evaluation set is called 'deterministic' (10 initial states per task), the paper does not specify whether the seeds that generate those initial states are fixed and documented. For a benchmark that aims to be reused, all random seeds that affect initial state sampling, perturbation realization, and rollout should be recorded and published. Without this, independent researchers cannot exactly reproduce the reported results. The authors should either confirm that all evaluation seeds are fixed and available, or provide seed lists for the 36 held-out tasks.

5 minor objections

The paper says, at Table 3

Pith says: minor objection 1 of 5

The PCR_S values for all policies (11.2–12.5%) are very close to the initial-state slotting baseline PCR_init_S = 12.31% (Table 14). This should be explicitly noted in the main text to clarify that these values largely reflect pre-existing valid destination workpieces, not successful manipulation. Currently this is mentioned in §4.2 but could be made more prominent.

The paper says, at §4.2, Table 3

Pith says: minor objection 2 of 5

The reported PCR_T for π0.5 is 0.4% overall, corresponding to approximately 1.4 episodes out of 360. The paper states '1.1% in S2D' (which is ~1.3 out of 120 rollouts). This is a very small number of successful target transfers. The authors should specify the absolute count (e.g., '1 episode' or '2 episodes') alongside the percentage to prevent overinterpretation.

show 3 more

The paper says, at Figure 4

Pith says: minor objection 3 of 5

Figure 4B reports PCR_T as a percentage out of 72 planned placements per condition. The bar chart shows per-step release-on-tray success. It would be helpful to also report the total number of trials that achieved at least one placement, not just the aggregate PCR_T.

more of objection 3

Also, the figure caption refers to 'release-on-tray success' but PCR_T counts all target transfers; clarify whether the step-level bars count workpieces successfully released on the tray (not necessarily fully slotted) or whether they correspond to the same PCR_T definition.

The paper says, at §3.2, Eq. (3)–(4)

Pith says: minor objection 4 of 5

The notation τ_src and τ_dst are used both as symbols and as text (e.g., 'τ_src'); the equations appear correct but the accompanying text should define d_src and s_slot explicitly in a table or immediately after the equations for clarity.

The paper says, at Appendix B.6.1

Pith says: minor objection 5 of 5

The paragraph on standards states that 'these standards provide reporting context but neither prescribe nor validate the simulator-domain bounds below, which are fixed by the benchmark generator.' This is an honest limitation. However, the paper uses ISO 9283 terminology (pose accuracy, repeatability) throughout the main text; a brief note in §3.4 or §A acknowledging that the implemented perturbation ranges are not validated against real physical measurements would help set reader expectations.

What would settle it

Equalize simulation frame counts between conditions: subsample LIBERO-Industry frames to match the 313,102 frames of S1, or upsample original-LIBERO frames to match the 620,660 frames of S2, then repeat the physical D2S evaluation (same robot, same 95 physical demonstrations). If the observed advantage of LIBERO-Industry co-training disappears or reverses, the original result is attributable to data quantity, not benchmark content.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 14 Oct 2026

  1. Reader holds up confidence moderate
  2. Skeptic partly agrees: says it holds up only with conditions 36 s · 25,790 tokens read, 4,294 written Sim-to-real co-training comparison confounds simulation data source with data quantity; the observed advantage of LIBERO-Industry simulation data may arise from larger frame count rather than benchmark content.
  3. Editor wrote the verdict above 10 s · 26,634 tokens read, 897 written
  4. Referee major revision: 2 major objections, 5 minor 27 s · 26,101 tokens read, 3,170 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper does something genuinely useful: it builds a benchmark that turns industrial process constraints—conveyor speeds, execution windows, geometric tolerances—into controllable variables for generalist robot policies. The task design is thoughtful, the progression from static to dynamic is principled, and the diagnostic stress analysis (even if not causal) gives a clear picture of where current policies fall apart. Credit where due: this is the first benchmark in this space that takes industrial constraints seriously as evaluation dimensions rather than just background noise.

Now the soft spots. The stress-test concern is real: the sim-to-real co-training comparison (S1 vs S2) does not hold frame count constant. S2 has roughly twice as many simulation frames as S1. That confound means you cannot tell whether the observed advantage comes from the industrial content of LIBERO-Industry or simply from more training data. The paper says corpus size is held constant (Appendix E.4), but it means number of demonstrations, not frames—and frames are what the model actually sees. This weakens the headline claim that co-training with LIBERO-Industry simulation data yields better transfer. It is not a fatal flaw, but the conclusion should be softened, or the experiment controlled properly.

Other issues are minor. The simulation results (0% SR across all policies) are unsurprising for such hard tasks, but the paper could be clearer about what baseline performance actually means. The real-world evaluation is limited to one robot and one D2S scenario. The paper acknowledges its own limitations (diagnostic stress, lack of confidence intervals), which is good.

Overall: this is a solid benchmark contribution. The data confound is real and needs fixing, but the benchmark design itself is valuable. I would bring it to reading group, and I would accept a peer-review invitation on it. This is a serious piece of work.

The strongest objection

The sim-to-real experiment in Section 4.4 compares two co-training conditions that differ in both the source of simulation data and its quantity. S1 uses 1,724 original-LIBERO demonstrations (313,102 frames) while S2 uses 1,412 LIBERO-Industry demonstrations (620,660 frames). Both employ a 9:1 simulation-to-real frame sampling ratio, so S2 injects roughly twice as many simulation frames per training step. The paper states that the comparison isolates the simulation data source and 'corpus size' (Appendix E.4), but corpus size in terms of frames is not held constant—only physical supervision is fixed. This confound means the higher PCR_T observed for S2 (34.7% vs 26.4% at 0.03 m/s; 23.6% vs 4.2% at 0.05 m/s) cannot be confidently attributed to the industrial content of LIBERO-Industry; it may simply reflect more simulation training data. As a result, the central claim that co-training with LIBERO-Industry simulation data yields better transfer is weakened, because the experiment does not control for a straightforward alternative explanation.

The reader's weakest_assumption focused on the realism of kinematic overrides in simulation. While that is a valid concern about external validity, the paper acknowledges it (Appendix B.6.1) and the sim-to-real experiment provides some empirical validation despite the overrides. However, a more immediate internal validity concern is the uncontrolled simulation data quantity in the co-training comparison, which undermines the support for the claim that LIBERO-Industry simulation data specifically improves physical transfer. Since the benchmark design and task suite remain valuable independently, I recommend a conditional verdict: the paper should either present a controlled comparison (equal frame counts) or explicitly qualify the co-training result as non-isolating. This does not reject the paper but asks for a clarification or additional experiment. My partial agreement reflects that the reader correctly identified simulation realism as a key assumption, but I find the data quantity confound to be a more direct threat to a central supporting result.

The referee's report

The benchmark addresses a clear gap: existing manipulation benchmarks abstract away industrial process constraints (timing, moving workpieces, tolerances). The task design is detailed and systematically varies constraints, which could drive future evaluation. The paper ships a large set of tasks (216 total, 36 held-out), code, and assets, and performs extensive simulation experiments with five distinct policy families. The real-world validation, while limited to D2S, provides an initial test of sim-to-real transfer. If the benchmark is adopted, it could meaningfully advance evaluation of robot policies for flexible manufacturing.

Pith's reading of the paper (a machine's summary, not the authors')

The paper argues that existing robot manipulation benchmarks do not capture the industrial process constraints—product variability, moving material flows, finite execution windows, and tight geometric tolerances—that real manufacturing lines impose. To fix this, the authors introduce LIBERO-Industry, a 216-task benchmark built on the LIBERO framework that translates these constraints into controllable simulation conditions across three scenarios: static-to-dynamic loading, dynamic-to-static unloading, and dynamic-to-dynamic transfer. When evaluated on the held-out tasks, five state-of-the-art generalist policies (including vision-language-action models) achieve zero full-task success, with success rates falling from 97% to 0% as constraints are combined in a stress progression. In real-world unloading trials on a physical robot, co-training with LIBERO-Industry simulation data yields higher target-transfer completion than co-training with the original LIBERO data, especially at faster conveyor speeds and under visual perturbations, suggesting that better-matched simulation data improves transfer.

What carries the argument. The argument's load-bearing mechanism is the benchmark's process-aware task generator, which conditions every task instance on material-flow dynamics. It uses a BDDL specification to define objects, initial states, transport conditions, and completion goals, with procedural layout sampling that includes feasibility and kinematic-safety gates. In simulation, two algorithms enforce conveyor transport and slot-arrest physics: Algorithm 1 applies kinematic anchoring to synchronize workpieces with the belt upon contact, and Algorithm 2 performs multi-stage velocity arrest during insertion, locking the object in place. These overrides bypass physics-engine chattering for rigid industrial constraints. The benchmark also defines episode horizons from transport schedules, enabling controlled time-pressure evaluation. The three difficulty levels—tray (Level 1), groove (Level 2), and form-fitting inlay (Level 3)—vary geometric precision, visual perturbations, and transport dynamics, allowing incremental stress diagnosis. Three metrics (success rate, target transfer PCR_T, and slot occupancy PCR_S) capture different aspects of task completion.

Core claim. On the paper's own terms, the central discovery is that industrial process constraints create systematic failure modes in current generalist robot policies that standard benchmarks do not expose. By controlling conveyor speed, execution window length, geometric precision, and object count independently, the benchmark reveals that success drops discontinuously when the task expands from one to multiple objects on moving conveyors (D5 stage, 2.1% success) and that no policy completes any of the 36 held-out tasks in full (0.0% SR). A secondary finding is that when fine-tuning for a physical dynamic-to-static unloading task, substituting LIBERO-Industry simulation data for original LIBERO data—while keeping the real demonstrations and training procedure identical—raises target-transfer completion from 26.4% to 34.7% at 0.03 m/s and from 4.2% to 23.6% at 0.05 m/s, with consistent improvements across five visual perturbations. The paper interprets this as evidence that simulation data that faithfully reflects the target operating conditions provides better transfer than general-purpose robot data.

The premise it rests on. The paper assumes that the simulation's kinematic overrides (Algorithms 1 and 2) for conveyor transport and slot-arrest physics are faithful enough to real-world contact that the relative ordering of policies and the improvement from co-training will generalize to other physical hardware and factory conditions; the authors acknowledge that these bounds are fixed by the benchmark generator rather than validated against physical measurements.

If the paper is right (Pith's inferences, not the authors')

Current vision-language-action and world-action models are not yet viable for production-line manipulation under typical industrial constraints, as shown by zero full-task success across all 36 held-out tasks.

The benchmark provides a standardized tool for tracking progress: future policies can be compared against the same 0% baseline and the partial-completion scores, making improvements measurable.

Co-training with process-constrained simulation data outperforms co-training with general-purpose simulation data in real-world transfer, suggesting that data relevance to the task's operating conditions is a key variable in fine-tuning for industrial applications.

The diagnostic stress progression pinpoints multi-object sequences and dynamic transport as the primary failure bottleneck, informing policy architecture or training curricula.

The finite execution windows, drawn from transport-arrival intervals, mean that policies must complete transfers within a pre-allocated time window, introducing a time-pressure dimension absent from most embodied-AI benchmarks.

The zero full-task success rate for all five policies implies that the distance between generalist robot research and industrial deployment may be much larger than standard benchmarks (LIBERO, CALVIN, Meta-World) suggest, because those benchmarks do not penalize slow execution or geometric imprecision to the same degree.

The stress progression shows that the addition of multi-object sequences on a moving conveyor (D5) is the most disruptive single change, collapsing success from 43.8% to 2.1%; this points to a specific weakness in handling temporal scheduling and repeated grasps under time pressure.

Since the simulation-kinematic overrides (Algorithms 1 and 2) are not validated against physical contact measurements, the benchmark's predictive value for other robot-hardware combinations or conveyor types remains an open question; a multi-embodiment replication would be needed to confirm the generality of the co-training advantage.

The fact that a simulation corpus with fewer total episodes (1,412 vs 1,724) but higher task match leads to better physical transfer suggests that dataset composition—specifically the inclusion of the target process constraints—matters more than dataset size for sim-to-real transfer in industrial manipulation.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Industrial manipulation in flexible manufacturing, e.g., high-mix, low-volume (HMLV) electronics assembly, requires robot policies to adapt to changing products and production conditions. Such settings impose industrial process constraints, including product variants, moving material flows, finite execution windows, and tight geometric tolerances. However, existing evaluation frameworks do not systematically capture these constraints when assessing generalist robot policies. To address this gap, we introduce LIBERO-Industry, a novel benchmark built on the widely used robot learning benchmark LIBERO, that translates the aforementioned industrial constraints into controllable task conditions and measurable success criteria. Specifically, LIBERO-Industry comprises three classic task scenarios: Static-to-Dynamic (S2D) loading, Dynamic-to-Static (D2S) unloading, and Dynamic-to-Dynamic (D2D) transfer. The suites contain 216 tasks in total that systematically vary object geometry, conveyor speed, interaction-window duration, and geometric acceptance region size. Extensive experiments across five generalist robot policies (including vision-language-action models and world-action models) show that performance progressively degrades as industrial constraints are introduced and combined. Under a shared fine-tuning and evaluation protocol, the LIBERO-Industry co-training condition achieved higher observed target-transfer completion than the original-LIBERO condition.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-14

The authors' main claim, as the classifier picked it out 0.95

“Extensive experiments across five generalist robot policies (including vision-language-action models and world-action models) show that performance progressively degrades as industrial constraints are introduced and combined.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.20 · states limitation 0.25 · human participants 0.03 · controlled comparison 0.94 · randomized assignment 0.03 · states quantitative result 0.88 · compares with prior results 0.92 · mentions code or data availability 0.04

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Industrial manipulation in flexible manufacturing, e.g., high-mix, low-volume (HMLV) electronics assembly, requires robot policies to adapt to changing products and production conditions. background 1.00 0.05 0.03 asserted 1.00
Such settings impose industrial process constraints, including product variants, moving material flows, finite execution windows, and tight geometric tolerances. background 1.00 0.07 0.02 asserted 1.00
However, existing evaluation frameworks do not systematically capture these constraints when assessing generalist robot policies. gap 1.00 0.06 0.03 asserted 1.00
To address this gap, we introduce LIBERO-Industry, a novel benchmark built on the widely used robot learning benchmark LIBERO, that translates the aforementioned industrial constraints into controllable task conditions and measurable success criteria. contribution 1.00 0.92 0.09 asserted 1.00
Specifically, LIBERO-Industry comprises three classic task scenarios: Static-to-Dynamic (S2D) loading, Dynamic-to-Static (D2S) unloading, and Dynamic-to-Dynamic (D2D) transfer. method 0.67 0.84 0.04 asserted 1.00
The suites contain 216 tasks in total that systematically vary object geometry, conveyor speed, interaction-window duration, and geometric acceptance region size. method 0.78 0.83 0.04 asserted 1.00
Extensive experiments across five generalist robot policies (including vision-language-action models and world-action models) show that performance progressively degrades as industrial constraints are introduced and combined. result 1.00 0.96 0.46 asserted 1.00
Under a shared fine-tuning and evaluation protocol, the LIBERO-Industry co-training condition achieved higher observed target-transfer completion than the original-LIBERO condition. result 1.00 0.96 0.29 asserted 1.00

What kind of paper the classifier reports this is: dataset or benchmark 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-14: 1 request, 6,343 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (14)
Figure 1 of the paper
Figure 1. LIBERO-Industry overview. Motivated by industrial process constraints, the benchmark evalu￾ates generalist robot policies across three classic industrial scenarios. production changes harder to accommodate with fixed systems [Koren et al., 1999]. Moving material flows impose finite time windows for manipulation [Islam et al., 2021], while parts must meet tight geometric tolerances despite manufacturing and perception errors [Jiang et al., 2022]. Traditional automation works reliably when tasks and environments remain stable, but frequent changes often require reprogramming,…
Figure 2 of the paper
Figure 2. Task generation and scripted demonstration collection in LIBERO-Industry. BDDL spec￾ifies objects, process states, and goals; sampled layouts pass a feasibility gate or are resampled (top). A task-specific scripted expert attempts a rollout from a fixed initial state and seed (bottom). The diagram shows completion, kinematic-safety, and replay checks before accepted trajectories are stored as HDF5 demonstrations with provenance. Failed attempts can prompt replay-guided collector refinement. Temporal opportunity condition. For the continuous source process, target workpieces…
Figure 3 of the paper
Figure 3. Cumulative diagnostic stress progression for π0.5. Each bar denotes complete-task success of the stress settings for n trials (D0–D1: n = 200; D2–D6: n = 240), and PCRT denotes mean target￾object transfer completion. Stars show mean policy steps over successful trajectories; D6 has no successful trajectories. The progression is diagnostic rather than causal. Figure 4B reports target-transfer completion at conveyor speeds of 0.03 and 0.05 m/s. At 0.03 m/s, S2 achieves PCRT = 34.7% (25/72), compared with 26.4% (19/72) for S1; at 0.05 m/s, the corresponding values are 23.6% (17/72) and…
Figure 4 of the paper
Figure 4. Real-world validation. A: AUBOi5H–CTAG2F90D D2S workcell. B: Per-step release-on-tray success for S1 and S2 at conveyor speeds of 0.03 and 0.05 m/s over 24 trials per condition; the table reports aggregate target-transfer completion as PCRT out of 72 planned placements. C: Target-transfer completion (PCRT) under five physical operating-condition variations, with exact values labeled. suites, with controlled variations in task configuration and operating conditions. Our results show perfor￾mance degradation across the cumulative industrial stress progression and higher D2S…
Figure 5 of the paper
Figure 5. Representative LIBERO-Industry task instantiations. Rows show Dynamic-to-Static (D2S),
Figure 6 of the paper
Figure 6. Representative LIBERO-Industry workpiece and fixture assets. From left to right, the panels
Figure 7 of the paper
Figure 7. Declared placement-region centers. The dashed
Figure 8 of the paper
Figure 8. maps each rollout to the first unmet manipulation phase, from pre-reach policy freeze and reach failure through post-release goal failure. OBSERVED TRAJECTORY goal region start reach grasp lift transport place REACH Target reached? EE-object distance < 50 mm GRASP Grasp acquired? env._check_grasp(...) = True LIFT Object lifted? ≥20 mm vertical or ≥30 mm XY TRANSPORT Grasp signal lost? grasped becomes False PLACE Task succeeded? after release inside target XY Policy froze before reach? EE drift ≤ 5 mm for 40 frames EARLY RELEASE Released inside goal? inside receptacle XY bounds GRASP…
Figure 9 of the paper
Figure 9. Representative failure trajectories. Each row shows three temporally ordered frames from one failed rollout, followed by the trajectory-level classifier output and its frequency within the corresponding diagnostic stage. The examples link the aggregate categories in
Figure 10 of the paper
Figure 10. Representative Lift-scene setup and wrist observations. The comparison checks robot–gripper
Figure 11 of the paper
Figure 11. Recovery pipeline. A candidate is added to the training dataset only after generation and replay
Figure 12 of the paper
Figure 12. Physical D2S workcell and fixtures. Left: AUBOi5H–CTAG2F90D workcell with conveyor, target tray, and marked HIKON camera. Right, reading order: Pico teleoperation equipment, example OOD objects, Level 2 grooves, and Level 3 inlays. 34
Figure 13 of the paper
Figure 13. Pico-assisted physical demonstration collection. Six photographs illustrate the collection pro￾cedure: initial scene, approach, grasp, transfer, placement, and release for a Level 1 tray transfer. 36
Figure 14 of the paper
Figure 14. Recorded physical demonstration id 001. Stored language: “Retrieve the items from the conveyor and set them onto the plate.” Each column pairs the table-camera view (top) with the wrist-camera view (bottom) at the same observation index of the aligned 10-Hz HDF5 episode. Times are relative to its first observation. Accepted dataset. The accepted physical dataset contains 95 demonstrations and 41,257 frames. The accepted counts for Levels 1, 2, and 3 are 32, 32, and 31, respectively. Episode id 092 was excluded because its nearest wrist-camera frame failed the predeclared 25 ms…

Figures are the authors', reproduced from the paper.

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment