Geometry Guided Self-Consistency for Physical AI
1. Paper overview
In one sentence: KeyStone selects a representative action chunk from parallel samples without retraining, gaining simulated task success when useful behaviors dominate and the GPU has spare sampling capacity. e02e03e04e08e09e10e11e13
| At a glance | What to know |
|---|---|
| Research problem | Source description A diffusion or flow-matching policy commits to one randomly generated action chunk, executes it open-loop, then observes again. A poor draw can derail an otherwise successful episode; repeated decisions create repeated opportunities for failure. The authors ask whether multiple candidates can improve reliability without training a judge or materially delaying each control round. e02e03e08 |
| Core mechanism | |
| A key reported result | GR00T N1.6 manipulation on SimplerEnv–WidowX: 63.3 ± 2.9% Task success rate, mean ± one across-repeat standard deviation. Simulation; K=4 versus K=1; task macro-average, five evaluation repeats 50.0 ± 4.3%; +13.3 percentage points Largest aggregate gain; an absolute improvement, despite the abstract's percent shorthand. e07e09 |
| Reading caution | Source description A confident incorrect majority can dominate selection and amplify failure; no selector can recover a successful chunk absent from its candidates. Batch-one latency results may not transfer to GPUs already serving multiple robots. e13 |
Core contributions
Figure 1. Repeated action sampling creates repeated opportunities for geometric consensus to influence an episode. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the upper timeline: arrows carry each inferred chunk into execution, followed by a later inference round. The illustrated action contains position, orientation and gripper components; this example is not a universal action layout for every tested policy. Below, each dot represents a complete chunk, projected separately at its round using PCA. The caption identifies GR00T N1.6 on SIMPLER and labels candidates by executing them to episode end. Teal points mark successes, orange points failures, and stars mark KeyStone selections. The connecting teal curve links decisions across rounds; it is not a single trajectory in a shared PCA coordinate system. e02e03e05e13
What it supports. The visual motivates choosing a representative sampled behavior: successful chunks concentrate in these examples, while failures spread away from them. Sections 3.1–3.3 turn that observation into full-chunk L2 distances, a guard and medoid selection. The selected output remains an actual model sample rather than an average of different behaviors.
Where the evidence stops. Outcome colors are retrospective diagnostic labels, unavailable to the deployed selector. Separate PCA projections and the illustrative inset do not prove that every dense region is successful; Section 5 explicitly permits an incorrect dominant cluster.
2. Motivation
2.1 The problem and the proposed response
A diffusion or flow-matching policy commits to one randomly generated action chunk, executes it open-loop, then observes again. A poor draw can derail an otherwise successful episode; repeated decisions create repeated opportunities for failure. The authors ask whether multiple candidates can improve reliability without training a judge or materially delaying each control round. e02e03e08
2.2 What this reading follows
Robot policies that generate action chunks by diffusion can respond differently to the same observation because each prediction starts from fresh noise. KeyStone turns that randomness into a small candidate set, then uses geometric agreement to choose what the controller executes. The useful distinction is between generating better candidates and choosing better among existing ones: this paper does the latter. Its efficiency argument rests on sharing expensive context computation while batching compact action trajectories. The illustrations below connect that mechanism to the strongest simulation result, the hardware-dependent sampling budget, a near-saturated policy, and the comparison with a learned selector. They also expose where the evidence stops. e02e03e04e08e09e10e11e13
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is VLA. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The foundational action-policy/planning placement is consistent with an inference wrapper applicable to existing VLAs and a WAM. KeyStone introduces neither a joint future/action predictor nor inverse dynamics, and no new integrated learned architecture. Architecture, prediction paradigm and quadrant are therefore appropriately not applicable to the wrapper itself; its host model's world embedding does not establish a One Model classification for KeyStone. e03e04e06
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Spend additional sampling on actions after computing context once
At a fixed control round, camera observations, robot state and instruction are identical for every candidate. KeyStone first lets the host model encode those inputs, then reuses the resulting features or cache while introducing independent noise for each action sample. This separates contextual work from stochastic action decoding. In the Fast-WAM integration, the video backbone runs once to provide a world embedding; the action decoder's denoising loop is the component batched across candidates. The wrapper does not introduce a separate predicted-video judge. Appendix A.1 describes logical expansion with shared storage and a single batched pass per denoising or flow update. The resulting candidates still pass through one selector and the original controller. Figure 2 determines how many parallel samples the tested hardware can accommodate economically. e03e04e06e08e14
5.2 Distinguish a representative sample from an averaged trajectory
After flattening complete action chunks, KeyStone builds one pairwise L2 distance matrix. Its global medoid is the sample with the smallest sum of distances to all others. The guard then measures mean–medoid separation relative to the median pairwise distance. Section 3.3 and Listing 2 agree on the branch: when s is below τ, default 0.3, return the global medoid; otherwise cluster the candidates. The default k-means step uses two clusters and returns the medoid of the largest one. This ordering matters: the mean diagnoses whether to split, but is never executed as a synthesized action chunk. A medoid preserves membership in the sampled set. It does not establish correctness, because a representative member of a confidently wrong cluster can still fail. e03e05e13e14
5.3 Separate evidence for success, selection and runtime
Reader analysis: the paper's strongest case combines three different comparisons rather than one universal guarantee. Table 1 measures task success for the complete wrapper at each model's selected candidate budget. Figure 2 supplies the runtime reason for those budgets, while Table 5 asks whether changing selector components alters gains. The cosine comparison supports using magnitude-sensitive geometry, but it does not test whether every successful behavior forms the largest cluster. Table 2 reveals both a small suite-level regression and an unresolved arithmetic discrepancy: the four ten-task suite means yield 97.5% and a 0.75-point mean gain, whereas the source prints 97.8% and +1.0 point Overall. These derived checks do not replace the author-reported results. Table 4 offers nearby but unequal point estimates against TACO without uncertainty. These distinctions keep the conclusion specific: geometric selection helps the evaluated simulation policies under the reported configurations. The next useful experiments should isolate the untested guard and measure whether the sampling budget survives realistic accelerator contention. e08e09e10e11e12e13
5.4 Training and inference
During training
KeyStone modifies no model weights, retrains no component and adds no optimization objective or training data. It presupposes a useful pretrained or fine-tuned base policy; the paper evaluates a wrapper rather than proposing a new learned world model. e06e13
During inference
Profile the deployment hardware before choosing K. On NVIDIA A6000-48GB GPUs, the paper uses K=4 for GR00T N1.6 and X-VLA, and K=16 for the other models. The baseline path is unchanged when disabled or K=1. Latency includes generation and selection; the reported regime assumes one robot with a dedicated GPU. e07e08e13
5.5 Implementation flow
- Encode once, sample in parallel
Reuse VLM features or transformer KV caches across candidates. For Fast-WAM, the video backbone produces one shared world embedding and the action decoder samples multiple chunks. Logical expansion with torch.expand shares context storage; each denoising or flow update processes the candidate batch together. e03e04e06
- Measure complete-chunk agreement
Flatten each ordered action chunk and compute the pairwise L2 distance matrix. The global medoid minimizes total distance to the candidates. No future-image scoring or learned success judge is introduced. e03e05
- Guard against an artificial split
Compare the distance between the sample mean and global medoid with the median pairwise distance. If the normalized score is below 0.3, return the global medoid. Otherwise run k-means with two clusters, random candidate initialization and at most ten iterations, stopping when assignments stabilize. e05
- Select and execute
Choose the largest cluster and return its within-cluster medoid. Because selection returns an existing sample, it avoids averaging distinct modes; it does not certify that sample as feasible or successful. The original controller executes the chunk and supplies subsequent observations. e02e05e06e13
6. Experiments & results
KeyStone improves stochastic robot policies by sampling several action chunks from one shared context and executing a geometrically representative candidate. It adds no learned selector or policy training. Its strongest reported simulation gain is 13.3 percentage points, while its latency advantage depends on spare GPU capacity. Consensus can also reinforce an incorrect dominant mode. The supporting mechanism, results and boundaries are traced in e02–e13.
The paper supplies a sampling/geometry schematic rather than a separate network architecture diagram, so Figure 1 anchors the method explanation alongside the equations and appendix. Its results concern simulated manipulation; no physical-robot success table is available. Table 5 ablates distance, sample count and cluster count, but supplies no separate guard ablation or uncertainty. Those absent experiment types cannot be illustrated from this source. e02e03e06e07e12
6.1 Read the original evidence
Figure 2. A useful candidate budget depends on the model's measured latency curve. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the horizontal coordinate as candidate count K, not denoising steps. The left vertical axis is logarithmic latency in milliseconds; the right is peak GPU memory in megabytes, also logarithmic. Match a model's color across the panels before comparing its cost. The latency includes sampling and selection in one action-generation round. GR00T's curve rises sharply at larger K and X-VLA also slows, so Section 4.2 chooses four candidates for those models. The other evaluated models use sixteen. This choice follows the curves rather than a fixed sampling multiplier for every policy; the right panel helps assess the additional memory burden. e04e07e08e13
What it supports. The curves support a practical operating region where several candidates cost little additional latency. Shared VLM context or world embeddings and batched action decoding explain the implementation strategy. They do not imply that arbitrary candidate counts are free: the rise for GR00T is visible within the plotted range.
Where the evidence stops. Measurements use A6000-48GB hardware and a dedicated-GPU, batch-one deployment setting. No exact zero-overhead bound or multi-robot guarantee follows from these curves; Section 5 requires profiling when accelerator load changes.
Table 1. Every author-reported workload-level mean improves, with the largest change on GR00T N1.6/WidowX. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as a separate paired comparison within one model and workload. The K column belongs to KeyStone, while the baseline always uses one sample. The two performance columns contain percentages, and the final delta is an absolute percentage-point difference. The caption specifies means and one standard deviation across five full runs, each macro-averaged across tasks. Section 4.1 explains that repeats resample policy noise; these deviations are not confidence intervals or training-seed variation. Compare GR00T across its two embodiments, then inspect Fast-WAM's separate RoboTwin row to see the scope of the WAM evidence. e06e07e08e09e10
What it supports. GR00T N1.6/WidowX improves from 50.0 ± 4.3% to 63.3 ± 2.9%, a 13.3-point gain at K=4. Fast-WAM improves from 90.0 ± 1.5% to 93.0 ± 1.0% at K=16. Together these rows establish gains in the tested VLA and WAM simulation integrations.
Where the evidence stops. Cross-row differences are not controlled architecture comparisons. The π0.5/LIBERO row retains the printed 97.8% KeyStone mean and +1.0-point gain, but Table 2's four ten-task suite means average to 97.5% and a 0.75-point gain. The stated task-macro averaging does not reconcile these values; the source discrepancy is unresolved. Variance reduction is not universal in the printed results: π0.5's standard deviation increases from 0.7 to 1.1, while X-VLA's remains 1.0. All listed workloads are simulated.
Table 2. The printed suite means do not reproduce the Overall row under the stated task-macro aggregation. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. This table expands the π0.5/LIBERO row rather than introducing another model. KeyStone uses K=16, and Spatial, Object, Goal and Long each contain ten tasks. Compare the suite rows with the printed Overall row rather than assuming they reconcile. Spatial rises from 96.0% to 98.0%, Object from 99.0% to 99.5%, and Long from 94.0% to 95.0%. Goal moves the other way, from 98.0% to 97.5%. The deltas are percentage points, and the paired numbers retain the source's reported standard deviations. Section 4.1 defines the workload mean as an average of per-task success rates. This is a performance breakdown, not an isolated mechanism ablation. e07e09e10
What it supports. Averaging the displayed four equal-size suites gives 96.75% for the baseline and 97.5% for KeyStone, a 0.75-percentage-point mean gain. Table 2 instead prints Overall values of 96.8 ± 0.7% and 97.8 ± 1.1%, with a +1.0-point gain repeated in its caption and discussion. The baseline mean rounds to 96.8%, but the KeyStone mean does not yield 97.8%. The report preserves the printed values and flags this unresolved inconsistency. The Goal decline also prevents any claim that consensus improves every task family.
Where the evidence stops. The derived 97.5% and 0.75-point gain are arithmetic checks of the displayed suite means, not replacements for unavailable raw experimental results. The source does not resolve the disagreement with its printed Overall row, and its aggregate standard deviation cannot be reconstructed from the suite standard deviations alone. The small changes and reported spread do not establish which suites benefit reliably. No paired significance test is provided, and saturation alone does not prove that the sampled action distribution is unimodal.
Table 4. KeyStone approaches the learned selector's point estimate without matching every task. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Keep this protocol separate from Table 1's Fast-WAM/RoboTwin row: the model here is π0.5. The source selects five tasks for which trained TACO selectors were available and evaluates each task fifty times. Both selection methods use K=16; the baseline uses one candidate. Compare TACO and KeyStone within each task before reading their macro averages. KeyStone is higher on Beat Block Hammer and Handover Block, while TACO is higher on Adjust Bottle, Move Can Pot and Place Object Stand. Unlike the aggregate table, these cells do not supply standard deviations alongside the success percentages. e06e11
What it supports. The evidence supports a low-training-cost alternative with a nearby macro point estimate: KeyStone reaches 64.4%, TACO 66.0%, and the baseline 58.0%. Avoiding a learned auxiliary selector is a real methodological distinction, but the reported means leave KeyStone 1.6 percentage points behind TACO in this comparison.
Where the evidence stops. The authors describe the methods as matching at the macro level. This selected five-task comparison supplies neither uncertainty nor an equivalence test, so it establishes neither statistical parity nor general replacement of learned selectors.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| GR00T N1.6 manipulation on SimplerEnv–WidowX Simulation; K=4 versus K=1; task macro-average, five evaluation repeats | 63.3 ± 2.9% Task success rate, mean ± one across-repeat standard deviation | 50.0 ± 4.3%; +13.3 percentage points Largest aggregate gain; an absolute improvement, despite the abstract's percent shorthand. e07e09 |
| GR00T N1.6 manipulation on SimplerEnv–Google Robot Simulation; K=4 versus K=1; task macro-average, five repeats | 86.7 ± 1.5% Task success rate, mean ± one standard deviation | 79.4 ± 3.5%; +7.3 percentage points Improvement extends to this second simulated embodiment. e07e09 |
| SmolVLA manipulation on LIBERO Simulation; K=16 versus K=1; task macro-average, five repeats | 57.2 ± 1.3% Task success rate, mean ± one standard deviation | 50.4 ± 2.1%; +6.8 percentage points Substantial remaining failure rate despite improvement. e07e09 |
| StarVLA manipulation on SimplerEnv–WidowX Simulation; K=16 versus K=1; task macro-average, five repeats | 59.7 ± 1.3% Task success rate, mean ± one standard deviation | 52.8 ± 1.6%; +6.9 percentage points Supports applicability to another action-generation backbone. e07e09 |
| X-VLA manipulation on SimplerEnv–WidowX Simulation; K=4 versus K=1; four-task macro-average, five repeats | 95.8 ± 1.0% Task success rate, mean ± one standard deviation | 92.7 ± 1.0%; +3.1 percentage points Table 3 localizes most benefit to less-saturated tasks; Put Carrot on Plate remains 100%. e07e09e10 |
| π0.5 manipulation on LIBERO Simulation; K=16 versus K=1; four suites of ten tasks, five repeats | 97.8 ± 1.1% Task success rate, mean ± one standard deviation | 96.8 ± 0.7%; +1.0 percentage point The value and comparison above preserve the author-reported Overall row. Table 2 gives four ten-task KeyStone suite means of 98.0%, 99.5%, 97.5% and 95.0%, whose equal-weight mean is 97.5%; the displayed baseline suite means average to 96.75%, giving a 0.75-percentage-point mean gain. These arithmetic checks do not reproduce the printed 97.8% Overall or +1.0-point gain under the stated task-macro aggregation. The discrepancy is unresolved; the derived values are not corrected raw experimental results. Goal declines from 98.0 ± 1.4% to 97.5 ± 0.7%, and the printed Overall standard deviation increases from 0.7 to 1.1, so neither taskwise improvement nor reduced aggregate variance is universal. e07e09e10 |
| Fast-WAM manipulation on RoboTwin 2.0 Simulation; K=16 versus K=1; task macro-average, five repeats | 93.0 ± 1.0% Task success rate, mean ± one standard deviation | 90.0 ± 1.5%; +3.0 percentage points Evidence for the tested WAM integration, not a survey of all WAM architectures or physical deployment. e06e07e09 |
| π0.5 comparison with TACO on RoboTwin 2.0 Five tasks with available trained TACO selectors; 50 evaluations per task; both selectors K=16 | 64.4% Macro-averaged task success rate | TACO 66.0%; K=1 baseline 58.0% Close point estimates without auxiliary training, but KeyStone trails TACO by 1.6 points. No uncertainty or equivalence test is supplied for Table 4. e11 |
| Selector ablation on GR00T N1.6/SimplerEnv–WidowX Simulation; vary one selector component; default L2, K=4, C=2 | Default +13.3 percentage points Absolute success-rate gain over K=1 | Cosine +0.5; doubled K +14.2; C=4 +8.3 Distance choice is consequential. Extra candidates give only a small gain here; Table 5 gives no uncertainty and does not ablate the guard. e12 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 5. Distance choice matters more consistently than adding samples or clusters in these tests. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Every numeric cell is a gain over that row's K=1 baseline, not an absolute success rate. Default means L2 distance, two clusters, and the hardware-profiled K: four for GR00T and X-VLA, sixteen otherwise. The next columns change one component at a time. On GR00T/WidowX, replacing L2 with cosine reduces the gain from 13.3 to 0.5 points; doubling K raises it to 14.2; four clusters gives 8.3. Read the numbers rather than the emphasis: the caption says bold marks column maxima, but the visible bold cells follow rowwise maxima. The crop preserves that source inconsistency. e05e08e12
What it supports. Cosine is worse than L2 on all five pairs, supporting the importance of action magnitude in this selector. Doubling K is not uniformly beneficial: X-VLA and GR00T/Google decline slightly. Increasing the cluster count also fails to produce a consistent improvement, making the default a practical choice for these workloads.
Where the evidence stops. No uncertainty is printed for the ablations. The table does not isolate the mean–medoid guard, coordinate scaling or centroid versus medoid choice, and it includes no Fast-WAM row. Its mechanism coverage is therefore narrower than the complete method.
7. Analysis & limitations
7.1 What the evidence leaves open
A confident incorrect majority can dominate selection and amplify failure; no selector can recover a successful chunk absent from its candidates. Batch-one latency results may not transfer to GPUs already serving multiple robots. e13
All reported benchmarks are simulated. Figure 1's per-round PCA illustrates correlation between geometry and outcome, rather than proving that Euclidean proximity guarantees equivalent contact behavior. Selection remains sensitive to the base policy's output representation. e02e03e07e12
Source inconsistencies require numerical reading: Section 3.3 says evaluated sample counts are below 16, whereas subsequent experiments use 16; Table 5's caption says bold marks column maxima, but the visible emphasis follows rowwise maxima. Table 2's four ten-task suite means average to 97.5% for KeyStone and 96.75% for the baseline, a 0.75-percentage-point mean gain, whereas its Overall row prints 97.8 ± 1.1% versus 96.8 ± 0.7% and +1.0 point. The stated task-macro averaging does not explain this discrepancy. The author-reported values are retained, and the arithmetic checks are not substituted for raw results; the inconsistency remains unresolved. e05e08e12e07e10
7.2 Questions for discussion
- When does the mean–medoid guard confuse a symmetric multimodal distribution with one compact mode?
- How much of the latency advantage survives multiple concurrent robot streams?
- Which action-coordinate scalings preserve the observed advantage of L2 over cosine?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the original checkpoints, testbed configurations and unchanged action post-processing. The source names A6000-48GB hardware, benchmark Docker images and PyTorch operations, but does not enumerate exact checkpoint revisions, container digests, software versions, per-model chunk lengths, denoising schedules, step budgets or every task's episode count. Typical counts are 24–100 per task; five repeats resample policy noise. e07e14
The appendix is an integration sketch: kmeans and expand_to_K remain helper calls. Action-coordinate normalization or weighting, k-means empty-cluster handling and random seeds are not specified. Proposed checks should freeze candidate batches to isolate selection, then separately profile shared-context sampling under increasing accelerator load. e03e05e08e13e14
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Hold candidates fixed to test the guard and distance rule
Reader-proposed check, not performed: on GR00T N1.6/SimplerEnv–WidowX, save identical K=4 candidate batches at cloned simulator states. Compare the full L2 selector with global-medoid-only, forced C=2 clustering without the guard, and the cosine variant. Keep action post-processing, continuation policy and evaluation seeds fixed. Record s, branch frequency, chosen candidate and eventual episode outcome; also evaluate each saved candidate from the cloned state to expose successful candidates that a selector misses. The guard's stated benefit is supported if it avoids harmful forced splits at low s. Repeated low-s selections from an unsuccessful mode despite available successful alternatives would reveal its failure boundary. e02e05e07e12e13e14
Check 2: Find where shared-context sampling stops preserving latency
Reader-proposed check, not performed: profile GR00T N1.6 and Fast-WAM on the reported A6000-48GB hardware at K=1,4,8,16, first with one robot stream and then with two and four concurrent streams. Compare shared-context expansion against a control that materializes separate context copies, keeping checkpoints, observations, chunk lengths and denoising schedules fixed. After warm-up and device synchronization, record median and tail per-round generation-plus-selection latency, peak memory and throughput. Predeclare an acceptable latency increase before choosing K. Loss of the single-stream operating budget under contention would confirm the paper's deployment caveat; unchanged budgets would extend the evidence to the tested concurrency levels. e04e07e08e13e14
8.3 Reading coverage
Visual audit: All 13 PDF pages were rendered and visually inspected, including the title/version/author block; Figure 1 and its caption; method Eqs. (1)–(6); evaluation setup, hardware and metrics; Figure 2; Tables 1–5; limitations; references; and both appendix listings. All six final crops were separately viewed with their labels, legends, axes and table rows intact. The s<τ branch in Section 3.3 was cross-checked against Listing 2. Figure 1's inference-to-execution arrows agree with its caption; its PCA panels are per-round projections. Table 5's boldface-caption inconsistency is disclosed, and numerical interpretations use the cells. The title page identifies v1 dated 9 May 2026, distinct from the catalog's submitted date of 8 May. No external supplementary material or repository was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract
- 1 Introduction
- 2 Background and Related Work
- 3 KeyStone
- 3.1 Method Overview
- 3.2 Latency-Preserving Parallel Sampling
- 3.3 Candidate Selection
- 3.4 Deployment Loop
- 4 Evaluation
- 4.1 Experimental Setup
- 4.2 Latency and Overhead
- 4.3 Task Success Rates
- 4.4 Ablations
- 5 Limitations
- 6 Conclusion
- References
- A Technical Appendices and Supplementary Material
- A.1 Implementation: KeyStone as an Inference-Time Wrapper, including Listings 1–2
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Identity: the title and all four authors match the catalog. The inspected title page identifies arXiv:2605.08638v1 [cs.RO], 9 May 2026; the catalog records submittedDate 2026-05-08. These dates are preserved as distinct metadata observations; no different revision or edition was supplied or compared.
- Acquisition omission preserved: text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This reading additionally inspected all 13 PDF pages, both figures and all five tables.
- Separate supplemental material availability has not been fully verified.
- The linked repository and external baseline resources were not inspected. Appendix listings were read as source text only; no code was executed and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block and arXiv margin
Observed title is Geometry Guided Self-Consistency for Physical AI. Authors are Yinwei Dai, Zhuofu Chen, Lijie Yang and Ravi Netravali; affiliation is Princeton University. The margin reads arXiv:2605.08638v1 [cs.RO], 9 May 2026.
Go to primary source ↓e02PDF p. 1, Section 1; PDF p. 2, Figure 1 panels (a)–(b), inset and caption
Episodes alternate sampling and open-loop chunk execution. Figure 1 labels GR00T N1.6/SIMPLER candidates by eventual episode outcome and projects chunks with per-round PCA. Green stars represent selected medoids; the inset shows example end-effector trajectories.
Go to primary source ↓e03PDF p. 4, Sections 3–3.3, Eqs. (1)–(3)
The wrapper samples K independent noises under shared encoded context, produces real candidate action chunks and flattens complete chunks for pairwise L2 distances. It selects an existing sample rather than a synthesized mean.
Go to primary source ↓e04PDF p. 4, Section 3.2; PDF p. 12, Appendix A.1, shared context expansion and Fast-WAM paragraphs
Candidates reuse VLM features, KV caches or world embeddings with torch.expand. Fast-WAM's video backbone runs once per round, followed by batched action-decoder denoising. Compact action sampling is argued to exploit memory-bandwidth-limited compute headroom.
Go to primary source ↓e05PDF p. 5, Section 3.3, Eqs. (4)–(6) and surrounding paragraphs
The mean–global-medoid score uses median pairwise distance in its denominator. Below τ=0.3 the global medoid is returned; otherwise k-means defaults to C=2, random candidate centroids, ten iterations maximum and assignment-stability stopping. The largest cluster supplies its medoid. The final paragraph says sample counts are below 16.
Go to primary source ↓e06PDF p. 5, Section 3.4; PDF pp. 12–13, Appendix A.1 and Listing 1
KeyStone intercepts action generation, leaves weights and downstream control unchanged, and requires no retraining. LeRobot, vla-eval and Fast-WAM integrations return the original action-chunk format after shared-context sampling and selection.
Go to primary source ↓e07PDF p. 5, Section 4.1 Testbed; PDF p. 6, Models, Workloads and Metrics
Evaluation uses NVIDIA A6000-48GB GPUs and simulated LIBERO, SimplerEnv and RoboTwin 2.0. Baseline configurations and processing are preserved. Typical per-task episode counts are 24–100; workload rates average per-task rates. Five full repeats resample sampling noise unless otherwise noted; reported deviations are across repeats.
Go to primary source ↓e08PDF p. 6, Figure 2, caption and Section 4.2; PDF p. 7, opening continuation
Plots show end-to-end generation-plus-selection latency and peak memory at K=1,4,8,16. Latency inflates beyond K=4 for GR00T and X-VLA; these use K=4 subsequently, while other models use K=16. The text qualifies the measurements by dedicated-GPU deployment and re-profiling for shared accelerators.
Go to primary source ↓e09PDF p. 7, Table 1, all seven model–workload rows and caption; Section 4.3
Author-reported baseline→KeyStone success means are π0.5/LIBERO 96.8→97.8, SmolVLA/LIBERO 50.4→57.2, GR00T/WidowX 50.0→63.3, X-VLA/WidowX 92.7→95.8, StarVLA/WidowX 52.8→59.7, GR00T/Google 79.4→86.7 and Fast-WAM/RoboTwin 90.0→93.0 percent. Table cells give across-five-run standard deviations and absolute percentage-point deltas. The π0.5/LIBERO aggregate is reproduced as printed, but its 97.8% KeyStone mean and +1.0-point gain are not recovered by averaging Table 2's four equal-size suite means; see e10.
Go to primary source ↓e10PDF p. 6, Section 4.1, Metrics; p. 8, Tables 2–3, all suite/task rows, captions and Table 2 discussion
Table 2 gives π0.5 LIBERO baseline→KeyStone means of Spatial 96.0±1.4→98.0±1.4, Object 99.0±1.4→99.5±0.7, Goal 98.0±1.4→97.5±0.7 and Long 94.0±1.4→95.0±1.4 percent; each suite contains ten tasks. Their displayed deltas are +2.0, +0.5, −0.5 and +1.0 percentage points. Reader arithmetic gives equal-weight means of 96.75% and 97.5%, and a mean gain of 0.75 point. The Overall row instead prints 96.8±0.7→97.8±1.1 and +1.0 point; the caption and discussion also claim the suite changes average to +1.0. Section 4.1 defines workload success as the mean of per-task rates and reports five evaluation-repeat means. The printed suite means therefore do not reproduce the aggregate under the stated aggregation. The cause is unresolved; 97.5% and 0.75 point are checks of displayed means, not corrected raw measurements, and no replacement Overall standard deviation is inferred. Table 3's separate X-VLA task deltas are Carrot 0.0, Eggplant +6.9, Spoon +1.4 and Cube +4.2; Carrot remains 100.0±0.0.
Go to primary source ↓e11PDF p. 8, Table 4, caption and Comparison with TACO paragraph
On five RoboTwin 2.0 tasks with released TACO selectors, π0.5 is evaluated 50 times per task. Baseline/TACO/KeyStone macro rates are 58.0/66.0/64.4 percent, both selectors at K=16. Task rows favor KeyStone on Beat Block Hammer and Handover Block, TACO on the other three. No uncertainty accompanies Table 4.
Go to primary source ↓e12PDF p. 8, Section 4.4 opening; PDF p. 9, Table 5, caption and ablation discussion
Ablations vary distance, sample count or cluster count on five pairs. GR00T/WidowX gains are default +13.3, cosine +0.5, doubled K +14.2 and C=4 +8.3 points. Cosine lowers gains on every pair; doubled K is mixed. No uncertainty or guard ablation is shown. The caption says columnwise bold maxima, although visible bold cells instead match row maxima.
Go to primary source ↓e13PDF p. 9, Section 5 Limitations
The method assumes useful candidate trajectories. It cannot recover success if all sampled candidates fail and can amplify a confidently incorrect majority. Latency measurements use batch size one; shared multi-robot accelerators may lose spare sampling headroom.
Go to primary source ↓e14PDF pp. 12–13, Appendix A.1, Selector implementation and Listings 1–2
The sketch uses PyTorch pairwise distances, median over off-diagonal pairs, s<tau, eps=1e-8, and medoid selection. expand_to_K and kmeans are helper calls rather than complete implementations. The appendix describes integration but does not enumerate software versions, model schedules or action-coordinate weighting.
Go to primary source ↓8.5 Primary sources
Geometry Guided Self-Consistency for Physical AI ↗
PDF · 6,965 extracted words
Source fingerprint
c4c1e06721d1339778a681fa70e9497bae5559e8dc10ce4174fad496e2fe2c28