acceptodds
Under review as a conference paper at ICLR 2027

Calibration in the Loop: Learning Early VLA Failure Detection with Condition-Wise Error Control under Deployment Shift

Abstract

A failure monitor for a vision-language-action (VLA) policy is a learned score and a threshold that carries the false-alarm guarantee. Under deployment shift the guarantee must hold per operating condition, and the per-condition rule places the deployed threshold at the heaviest successful tail, where a score trained by cross-entropy alarms late, after a retry can no longer succeed. We frame monitor training under deployment shift as a constrained optimization problem: maximize early detection of failures subject to a false-alarm tolerance under every deployment condition, and we solve it with calibration in the loop (CIL), a training framework that computes the calibration rule that will set the deployed threshold inside the objective. The rule is an order statistic of reference successes per condition and the maximum over conditions, exact and differentiable almost everywhere, so the score is trained under the threshold it will be deployed with while dual ascent enforces one constraint per condition; the loss, the constraints and the rule can be exchanged. One trained score serves a marginal and a high-confidence deployment mode whose worst-condition bound holds with probability . On held-out banks under the LIBERO-Plus catalog, CIL raises early detection within the tolerance from 39.3% to 59.2% for on LIBERO-Spatial and to 85.5% for OpenVLA-OFT when applied to a SAFE network, against detection baselines thresholded by the same rule; in a reset-and-retry loop the earlier alarms become completed tasks as the reset cost grows; and on a real xArm6 the monitor raises early detection from 34.9% to 44.4% within the tolerance and completes more tasks under recovery than the CoRe and TIDE monitors. Code and data recipes are available at https://anonymous.4open.science/r/CIL-4340/README.md.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Vision-Language-Action Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_oWBTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors proposed calibration in the loop method for monitoring VLA failures, and their framework computes the calibration rule that set the deployed threshold inside the objective. Their experiments showed relatively high performance on LIBERO-Plus, LIBERO-Spatial and other datasets.

Critical strengths

Thanks the authors for submitting to ICLR 2027! I think monitoring VLA’s failure, alerting and improving this loop is important. The authors also contributed a concrete step to this problem. They proposed a calibration in the loop method that computes the calibration rule that set the deployed threshold inside the objective.

  1. Well-motivated problem and well-organized calibration in the loop method.
  2. Extensive experiments that show the effectiveness of their calibration in the loop method.
Critical weaknesses

I did not see critical weakness. There are a few minor comments that I think the authors could take to improve their paper.

(Minor) The VLA operation figure in the Figure 2 should be clearer.

The authors need to explain tests A and B in L346 (and also appeared in L362, etc.). I found the descriptions of tests A and B in the appendix but it is currently confusing in the main text.

I think the authors need to briefly clarify what statistic the false alarm’s definition is depending on L316-317, “a success is a false alarm the first time its statistics crosses the threshold”

I think the authors at least need to report the numbers in the main text and their conclusions. Most of the results of the current 5.2 seems to be pointing to the appendix, which is difficult to read.

Questions for authors

NA. See the weaknesses session and consider improve these minor points.

2 comments

We thank Reviewer oWBT for the encouraging assessment and the careful reading. We appreciate the reviewer’s recognition that the problem is well motivated and that calibration in the loop is a concrete step toward reliable VLA failure monitoring.

W1 — VLA panel of Figure 2

Reply: Panel (a) will be redrawn in the revision. It will show the frozen VLA, the context vector xtx_t read from its internal state at each policy call, the per-call surprise sts_t, its running mean sˉt\bar{s}_t, and the crossing of the deployed threshold τ\tau that triggers the reset.

W2 — tests A and B

Reply: Tests A and B are two held-out test sets of π0.5\pi_{0.5} on LIBERO-Spatial, separate from the main test set used in Table 1, run at initial states 2–3 while the thresholds are calibrated at initial states 0–1. Test B is used for the recovery experiments of Sec. 5.4: on top of the LIBERO-Plus condition, each episode receives one transient disturbance that we add (the target object is pushed by 4–7 cm, or the gripper is forced open). Test A holds the same episodes without this disturbance, which tells how many of the failures in test B are rescuable by a reset. Sec. 5.1 of the revision will define both.

W3 — the statistic behind a false alarm

Reply: It is the episode statistic of Sec. 3, M(θ)=max⁡tsˉt(θ)M(\theta) = \max_t \bar{s}_t(\theta): the running mean sˉt\bar{s}_t of the per-call surprise st=softplus(θ⊤[xt;1])s_t = \mathrm{softplus}(\theta^\top [x_t; 1]), maximized over the eligible calls. A successful episode is a false alarm when M>τM > \tau, counted at the first call where sˉt\bar{s}_t crosses the deployed threshold. This follows the definition of SAFE (Gu et al., 2025): a successful rollout becomes a false positive whenever its score exceeds the threshold, and the detection time is the first such crossing. We will state it clearly in the camera ready version.

W4 — numbers and conclusions of Sec. 5.2 in the main text

Reply: Sec. 5.2 will state each result with its conclusion. (a) Training under the rule raises early detection over the same head trained by cross-entropy, with every condition within the tolerance: from 39.3% to 57.1% (marginal) and from 32.5% to 52.4% (high-confidence) on LIBERO-Spatial, and from 33.4% to 46.5% and from 24.9% to 34.7% on LIBERO-Long. (b) Differentiating through the calibration rule adds 5.5 to 10.8 points of early detection on top of direct constrained optimization (the stop-gradient control of Table 6). (c) Against SAFE-LSTM, the strongest baseline, CIL leads in both matched modes on both OpenVLA-OFT benchmarks: 85.5 vs 77.6 (marginal) and 69.4 vs 54.1 (high-confidence) on LIBERO-Spatial, and 67.2 vs 63.7 and 51.4 vs 34.2 on LIBERO-Long.

References: Gu et al., 2025 (https://arxiv.org/abs/2506.09937).

Dear Reviewer oWBT,

We hope we’ve addressed all your concerns. Since the discussion period ends in <24 hours, please let us know if you need any further clarification. If no further clarifications are needed, may we kindly request that you revise your scores upwards?

Thank you, The Authors

Reviewer_kLs7OpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Existing failure monitors typically train a score to distinguish successful and failed trajectories using cross-entropy, then calibrate a detection threshold afterward. Under distribution shift, the authors argue that this procedure can lead to late failure detection, especially when the deployed threshold must satisfy a false-alarm tolerance under every operating condition. CIL addresses this by optimizing early failure detection while differentiating through the condition-wise calibration threshold.

Critical strengths

The paper identifies an important mismatch between how failure detectors are trained and how they are deployed. Optimizing average classification accuracy or ROC-AUC does not necessarily yield effective early detection at a strict false-alarm threshold.

Critical weaknesses
  • Differentiating through calibration thresholds and optimizing under group-wise constraints have both been explored in prior work. The proposed method mainly combines these existing techniques for VLA failure detection. Its methodological novelty could be better justified.

  • Limited robustness to distribution shift. The theoretical guarantees assume calibration and deployment data follow the same distribution within each condition. However, Table 7 shows that even changing initial states can violate the 10% false-alarm tolerance (up to 13.6%). This weakens the paper's central claim of robustness under deployment shift.

Questions for authors
  • Since the false-alarm guarantee fails under initial-state shifts in Table 7, what types of deployment shift can CIL actually handle?

  • Why does the substantial improvement in early detection translate into relatively small gains in task completion? Could directly optimizing recovery success be more effective?

  • Since the method already uses condition labels, why not deploy separate thresholds for different conditions instead of using the maximum threshold across all conditions?

  • What is the key technical innovation beyond combining differentiable conformal calibration with existing constrained optimization methods?

5 comments
AuthorsOpenReview ↗

Response to Reviewer kLs7 (1/2)

We thank Reviewer kLs7 for the thoughtful review. We appreciate the reviewer's recognition that training a failure detector for average accuracy or ROC-AUC does not give early detection at a strict false-alarm threshold, which is the mismatch this paper targets.

W1 + Q4 — key technical innovation

Reply: To our knowledge, CIL is the first to put calibration in the training loop of a VLA failure detector. Existing VLA failure detectors, including the baselines of Table 1, train a score first and calibrate its threshold afterwards (SAFE, Gu et al., 2025; SAFECAST, Rajaprakash et al., 2026; ActProbe, Huang et al., 2026), so training never sees the threshold that deployment will use. CIL computes that threshold inside the objective with the exact rule used at deployment, each condition's order statistic and their maximum (Eq. 3), so the score is optimized at its deployed operating point. This lets us optimize any loss while preserving the false-alarm constraint we want: the guarantee comes from the final calibration of the frozen score and holds whatever loss was used in training (Propositions 1–2), so the early-detection loss and the constraints of Eq. 5 can be exchanged for other objectives (Sec. 4). The gradient update through the calibration rule brings a substantial gain on its own. To isolate it, the stop-gradient control of Table 6 removes only that gradient: it optimizes early detection directly under the same per-condition false-alarm constraints, with the same loss and dual ascent as CIL (Eqs. 5–6). The table places it between cross-entropy and CIL on the two held-out sets of π0.5\pi_{0.5}/LIBERO-Spatial: the detection test set (Table 1) and the recovery test set (the 1,272 disturbed episodes of the closed-loop study of Sec. 5.4, “test B” in Table 7). Differentiating through the calibration rule adds 5.5 to 10.8 points of early detection on top of direct constrained optimization.

Early detection (%)Cross-entropy (θ0\theta_0)Direct constrained (no gradient through τ\tau)CILGain from the gradient
Detection test set, marginal39.350.857.1+6.3
Detection test set, high-confidence32.544.052.4+8.4
Recovery test set, marginal28.531.442.2+10.8
Recovery test set, high-confidence19.025.631.1+5.5

W2 + Q1 — which deployment shifts CIL handles

Reply: The marginal mode holds the tolerance on average, so a single calibration draw can exceed it. Its bound is on each condition's false-alarm rate averaged over calibration draws (Proposition 1), and with a limited number of calibration successes per condition one draw can land above 10%. Even without any shift, over random within-condition splits the mean worst-condition rate is 8.9%, yet the worst condition exceeds 10% on 35.8% of splits (38.8% for the cross-entropy head; App. D). The 13.6% of Table 7 is a marginal-mode result of this kind, and the cross-entropy head reaches 12.9% under the same rule. The high-confidence mode takes the sample size into account: its binomial rank bounds the worst condition with probability at least 1−δ1 - \delta given the number of calibration successes (Proposition 2). It holds on 97.0% of the same splits, and under the initial-state shift of Table 7 it keeps the worst-condition rate at 9.2% (test A) and 6.8% (test B) (App. D).

In both modes, the guarantee covers every represented condition and any mixture of them in unknown proportions: since τ≥τc\tau \ge \tau_c for every cc, a mixture's false-alarm rate is at most the worst condition's, while a pooled threshold covers only the calibration mix. It also extends to an unseen condition whose successful runs score no higher than those of some represented condition, because the deployed threshold is set by the heaviest represented tail. Table 9 tests this by leaving each condition out of calibration: in the high-confidence mode every left-out condition stays within the tolerance (at most 8.1%), whereas a pooled threshold reaches 14.2–33.5% on four of them.

AuthorsOpenReview ↗

Response to Reviewer kLs7 (2/2)

Q2 — detection gains versus completion gains

Reply: Completion has little room at a cheap reset, and CIL takes most of it. On the recovery test set, 70.8% of the episodes complete with no monitor and 270 of the 372 failures are rescuable (their undisturbed twin succeeds), so a monitor that rescued every one of them and reset no success would complete 92.0%; CIL-M completes 89.1%, 86% of that gain. At a 20-step reset a late alarm still leaves time to retry, so earliness buys little; it pays once a reset costs time: at 140 steps CIL-M completes 83.0% against 75.6% for the calibration-only monitor (+7.4 [5.4, 9.4], Table 13), at 6.4% unnecessary resets against 63% and 75% for CoRe and TIDE. The table follows the 50 failures that CIL-M flags early and the cross-entropy head (θ0\theta_0-M) does not. At a 20-step reset the later alarm of θ0\theta_0-M still rescues most of them, and the two monitors break even (3 against 3); at 140 steps CIL-M alone completes 9 of them and θ0\theta_0-M alone none.

Failures flagged early by CIL-M and not by θ0\theta_0-M (n = 50)Reset cost 20Reset cost 140
Completed under both monitors2721
Completed only under CIL-M39
Completed only under θ0\theta_0-M30
Completed under neither1720

Directly optimizing recovery success leaves little to gain here: at a 20-step reset, even a monitor that rescued every rescuable failure would add only 2.9 points over CIL-M (92.0% against 89.1%). We leave a recovery-aware objective to future work, and CIL already supports it: the framework can optimize any loss under the same false-alarm guarantee, so the early-detection term of Eq. 5 can be replaced by a recovery objective (Sec. 4).

Q3 — separate thresholds per condition

Reply: Condition-specific thresholds need the condition at run time. In our setting the condition is not observed at deployment and the deployment environment may mix conditions which is the case in many real world environments; the labels only exist in the calibration data. CondMax is a single threshold that needs no label at run time, and its high-confidence bound on the worst condition needs no union bound over conditions (Proposition 2). With this threshold, the guarantee covers every represented condition and any mixture of them in unknown proportions: since τ≥τc\tau \ge \tau_c for every cc, a mixture's false-alarm rate is at most the worst condition's, while a pooled threshold covers only the calibration mix. It also extends to an unseen condition whose successful runs score no higher than those of some represented condition, because the deployed threshold is set by the heaviest represented tail. Table 9 tests this by leaving each condition out of calibration: in the high-confidence mode every left-out condition stays within the tolerance (at most 8.1%), whereas a pooled threshold reaches 14.2–33.5% on four of them. Sec. 3 will state this explicitly.

References: Gu et al., 2025 (https://arxiv.org/abs/2506.09937); Rajaprakash et al., 2026 (https://arxiv.org/abs/2608.04246); Huang et al., 2026 (https://arxiv.org/abs/2606.08508).

Dear Reviewer kLs7,

We hope we’ve addressed all your concerns. Since the discussion period ends in <24 hours, please let us know if you need any further clarification. If no further clarifications are needed, may we kindly request that you revise your scores upwards?

Thank you, The Authors

Reviewer_vMZhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Calibration-in-the-Loop (CIL), a framework that incorporates condition-wise conformal calibration into the training objective of VLA failure monitors. By differentiating through the CondMax threshold and optimizing early detection under per-condition false-alarm constraints, CIL aims to improve detection at the deployed operating point. The method is evaluated on LIBERO benchmarks, simulated recovery tasks, and a real robot.

Critical strengths
  1. Clear motivation. The mismatch between conventional classification objectives and calibrated deployment thresholds is a meaningful problem for early failure detection.

  2. Reasonable technical contribution. Incorporating the calibration threshold into training provides a principled approach to optimizing detection under false-alarm constraints.

  3. Extensive evaluation. The experiments cover multiple VLA policies, detector architectures, and deployment conditions, with useful ablations and supplementary analyses.

Critical weaknesses

W1. Limited robustness to unseen deployment shifts.

CIL relies on predefined operating conditions and representative calibration data. While it addresses differences across known conditions, it does not resolve calibration validity under unseen conditions or distribution shifts within a condition. This limits the generality of its deployment-robustness claims.

W2. The necessity and advantages of the proposed calibration-aware optimization are not fully established.

CIL improves detection under CondMax but retains its conservative global-threshold mechanism. It remains unclear how much of the improvement is attributable to differentiating through calibration rather than directly optimizing early detection under FPR constraints, or whether similar performance could be achieved using condition-specific calibration when condition labels are available.

W3. Experimental comparisons and validation protocols require stronger justification.

The main results involve different calibration modes and evaluation-bank-informed model selection, making it difficult to isolate the gains attributable to CIL. In addition, the real-robot recovery evaluation is insufficiently specified to establish the claimed closed-loop benefits.

Questions for authors
  1. Model selection (Section 5.1, Lines 312f; Appendix B, Lines 643f): The reported CIL configurations were selected among variants evaluated on the Fresh Bank. How many configurations were considered? Can the authors report results for models selected exclusively using development data?

  2. Baseline fairness (Table 1, Lines 286ff; Table 18, Lines 1442f): On OpenVLA-OFT/LIBERO-Spatial, SAFE-LSTM achieves 77.6% early detection with 3.6% worst FPR under marginal CondMax (Table 18), whereas Table 1 reports its high-confidence result of 54.1% against CIL-M's 85.5%. Why are different calibration modes compared? Can the authors provide matched M-vs-M and HC-vs-HC results?

  3. Real-robot evaluation (Table 4; Lines 453f; Appendix J, Lines 1215f): Table 4 reports recovery completion rates, while Appendix J states that no monitor ran on the physical robot. How were these recovery outcomes obtained? Were the retries physically executed or evaluated offline?

5 comments
AuthorsOpenReview ↗

Response to Reviewer vMZh (1/3)

We thank Reviewer vMZh for the careful reading. We appreciate the reviewer's recognition that the mismatch between classification objectives and calibrated deployment thresholds is a meaningful problem, and that placing the calibration threshold inside training is a principled way to address it.

W1 — unseen deployment shifts

Reply: The guarantee covers every represented condition and any mixture of them in unknown proportions: since τ≥τc\tau \ge \tau_c for every cc, a mixture's false-alarm rate is at most the worst condition's, while a pooled threshold covers only the calibration mix. It also extends to an unseen condition whose successful runs score no higher than those of some represented condition, because the deployed threshold is set by the heaviest represented tail. Table 9 tests this by leaving each condition out of calibration: in the high-confidence mode every left-out condition stays within the tolerance (at most 8.1%), whereas a pooled threshold reaches 14.2–33.5% on four of them. Under an initial-state shift inside the calibrated conditions (tests A and B), the margin of the high-confidence rank keeps the worst-condition rate at 9.2% and 6.8% (App. D).

W2 — source of the gain, condition-specific thresholds

Reply: The stop-gradient control of Table 6 is this baseline. It optimizes early detection directly under the per-condition false-alarm constraints, with the same loss and dual ascent as CIL (Eqs. 5–6); only the gradient through τ(θ)\tau(\theta) is removed. The table places it between cross-entropy and CIL on the two held-out sets of π0.5\pi_{0.5}/LIBERO-Spatial: the detection test set (Table 1) and the recovery test set (the 1,272 disturbed episodes of the closed-loop study of Sec. 5.4, “test B” in Table 7). Differentiating through the calibration rule adds 5.5 to 10.8 points of early detection on top of direct constrained optimization; on the detection test set the paired 95% intervals of this gain are [0.0, 12.2] (marginal) and [2.4, 14.1] (high-confidence).

Early detection (%)Cross-entropy (θ0\theta_0)Direct constrained (no gradient through τ\tau)CILGain from the gradient
Detection test set, marginal39.350.857.1+6.3
Detection test set, high-confidence32.544.052.4+8.4
Recovery test set, marginal28.531.442.2+10.8
Recovery test set, high-confidence19.025.631.1+5.5

Condition-specific thresholds need the condition at run time. In our setting the condition is not observed at deployment and the deployment environment may mix conditions which is the case in many real world environments; the labels only exist in the calibration data. CondMax is a single threshold that needs no label at run time, and its high-confidence bound on the worst condition needs no union bound over conditions (Proposition 2). Sec. 3 will state this explicitly.

AuthorsOpenReview ↗

Response to Reviewer vMZh (2/3)

W3 + Q1 — model selection

Reply: The development-only result is the paper configuration of Table 17. Every CIL model below has its checkpoint chosen on the development split, with no access to the detection test set. To isolate what CIL adds, the table pairs each score with the same score trained with CIL under the same threshold rule: early detection, with the worst-condition false-alarm rate in parentheses. The linear head gains 9.8 to 19.9 points with every condition within the tolerance, and SAFE-MLP gains 1.1 to 15.6 points; SAFE-LSTM gains 6.7 points on LIBERO-Long in the high-confidence mode.

π0.5\pi_{0.5} detection test set: early detection % (worst-condition FPR %)Without CILWith CILGain
LIBERO-Spatial
Linear head, marginal39.3 (7.2)57.1 (8.6)+17.8
Linear head, high-confidence32.5 (6.2)52.4 (8.1)+19.9
SAFE-MLP, marginal55.0 (12.4)60.2 (10.5)+5.2
SAFE-MLP, high-confidence47.6 (8.6)48.7 (9.1)+1.1
SAFE-LSTM, marginal70.2 (12.0)69.6 (13.9)−0.6
SAFE-LSTM, high-confidence61.8 (9.6)61.8 (9.6)0.0
LIBERO-Long
Linear head, marginal33.4 (4.4)46.5 (9.6)+13.1
Linear head, high-confidence24.9 (1.7)34.7 (4.6)+9.8
SAFE-MLP, marginal22.4 (9.3)38.0 (9.6)+15.6
SAFE-MLP, high-confidence20.8 (7.8)22.9 (6.4)+2.1
SAFE-LSTM, marginal49.4 (9.6)54.0 (10.7)+4.6
SAFE-LSTM, high-confidence39.1 (6.4)45.8 (5.0)+6.7

W3 + Q2 — matched M-vs-M and HC-vs-HC

Reply: In Table 1 every baseline is reported in the high-confidence mode, because that is the mode in which all baselines keep the worst-condition false-alarm rate within the 10% tolerance on all four benchmarks. In the marginal mode SAFE-MLP, SAFE-LSTM and ActProbe exceed it on π0.5\pi_{0.5}/LIBERO-Spatial (12.0–12.4%).

The matched comparison is below: early detection (%), with the worst-condition false-alarm rate in parentheses († above the tolerance). CIL leads SAFE-LSTM in both matched modes on both benchmarks: 85.5 vs 77.6 (M) and 69.4 vs 54.1 (HC) on LIBERO-Spatial, and 67.2 vs 63.7 (M) and 51.4 vs 34.2 (HC) on LIBERO-Long. The controlled comparison, each score with and without CIL under the same mode, is in the table of W3 + Q1 above. Table 1 will report both modes for every method.

Marginal (M vs M)π0.5\pi_{0.5} Spatialπ0.5\pi_{0.5} LongOFT SpatialOFT Long
kNN distance7.9 (4.9)5.7 (8.1)31.4 (6.4)20.5 (7.9)
SAFE-MLP55.0 (12.4†)22.4 (9.3)36.1 (6.4)29.3 (7.1)
SAFE-LSTM70.2 (12.0†)49.4 (9.6)77.6 (3.6)63.7 (8.9)
ActProbe57.6 (12.4†)41.1 (7.7)28.6 (4.5)14.3 (6.4)
CIL-M (ours)59.2 (9.7)46.5 (9.6)85.5 (5.5)67.2 (6.0)
High-confidence (HC vs HC)π0.5\pi_{0.5} Spatialπ0.5\pi_{0.5} LongOFT SpatialOFT Long
kNN distance3.1 (3.4)3.6 (4.4)10.2 (3.6)11.8 (4.7)
SAFE-MLP47.6 (8.6)20.8 (7.8)0.0 (1.8)0.0 (2.2)
SAFE-LSTM61.8 (9.6)39.1 (6.4)54.1 (1.8)34.2 (1.6)
ActProbe49.2 (9.6)29.3 (3.9)4.3 (0.9)2.6 (0.5)
CIL-HC (ours)61.8 (9.6)45.8 (5.0)69.4 (1.8)51.4 (1.2)
AuthorsOpenReview ↗

Response to Reviewer vMZh (3/3)

W3 + Q3 — real-robot recovery (Table 4)

Reply: The retries of Table 4 were executed physically on the xArm6, with each monitor running on the robot. The sentence in Appendix J that no monitor runs on the robot refers to the detection study of Table 3, where the monitors are trained and scored offline on the logged episodes. In the recovery runs a workstation with two RTX 4090 GPUs runs the policy on one GPU and the monitor on the other, so the monitor scores every policy call online from the latent returned by the same forward pass (69 ms per call with or without it, App. J), with control at 10 Hz. When it alarms, the arm stops, the scene is reset by a scripted return to the start pose, and the policy retries, up to three times within the 600-control-tick budget of an episode (App. J). Each monitor is evaluated on the three tasks, each on six object layouts under the three conditions (Fig. 4), and ± is the standard deviation over layouts. The platform is a UFACTORY xArm6 with the xArm gripper and two Intel RealSense D435i cameras (base and wrist, 640×480 at 30 fps), under a safety layer that limits the joint step to 0.30 rad per tick and the joint speed to 2.0 rad/s. Appendix J will carry this protocol.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.