pith. sign in

arxiv: 2606.06660 · v1 · pith:NNWS2BL2new · submitted 2026-06-04 · 💻 cs.AI · cs.PF· cs.RO

AEGIS: A Backup Reflex for Physical AI

Josef Chen This is my paper

Pith reviewed 2026-06-28 00:52 UTC · model grok-4.3

classification 💻 cs.AI cs.PFcs.RO
keywords robot manipulationpolicy escalationactivation probefailure detectionselective switchingLIBERO benchmark
0
0 comments X

The pith

A selective escalation method using activation probes on weak robot policies recovers 10.1% of lost trajectories by switching to stronger policies only at flagged high-risk steps.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Robot manipulation policies often degrade gradually after one bad step, with detectable signals before the failure becomes irreversible. AEGIS trains a lightweight probe on the frozen activations of a weak policy to identify these high-risk steps while lead time remains. Upon a positive flag the system hands control to a separate stronger policy for that step only. On the LIBERO-Spatial benchmark this recovers 10.1 percent of the trajectories the weak policy alone loses, compared with 4.6 percent for budget-matched blind escalation and 5.1 percent for random triggering. The advantage holds under pre-registered paired statistical tests and occurs while the stronger policy runs on just 38 percent of steps.

Core claim

AEGIS (Activation-probe Early-warning, Gated Inference Switching) places a probe on the activations of a frozen weak policy to detect high-risk steps with usable lead time and gates inference switching to a stronger separate policy only on those steps. On LIBERO-Spatial the method recovers 10.1 percent of trajectories the weak policy loses, versus 4.6 percent for budget-matched blind escalation and 5.1 percent for a random-trigger placebo, with significance under one-sided exact paired McNemar tests (Holm-Bonferroni adjusted) and an early-window AUROC of 0.764 read from the first 30 percent of steps.

What carries the argument

The activation probe on frozen weak-policy activations that flags high-risk steps to gate selective handoff to a stronger policy.

If this is right

  • The stronger policy need run on only 38 percent of steps, so the performance gain comes from timing rather than total compute.
  • No retraining of either base policy is required; only the probe is trained on activation data from the weak policy.
  • The conditional recovered-task-rate estimand isolates the benefit of timely handoff from overall policy strength.
  • The method preserves the weak policy's behavior on the majority of steps while protecting against the subset that would otherwise cause irrecoverable failure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the probe transfers across robot embodiments or task distributions, AEGIS could function as a modular safety reflex for any base policy.
  • The same early-warning idea might extend to non-robot domains where model outputs degrade gradually, such as long-horizon planning sequences.
  • Replacing the fixed stronger policy with a family of increasingly capable models could allow graduated escalation based on probe score.

Load-bearing premise

The probe can detect high-risk steps with enough lead time for the stronger policy to intervene before the state becomes irrecoverable.

What would settle it

A replication in which the probe's early-window AUROC on the first 30 percent of steps falls below 0.65 and the paired recovery-rate difference versus blind escalation loses significance under the pre-registered McNemar tests.

Figures

Figures reproduced from arXiv: 2606.06660 by Josef Chen.

Figure 1
Figure 1. Figure 1: Why timing is the whole problem. Schematic phase portrait of long-horizon manipulation: under the weak policy alone, a perturbed trajectory spirals inward and compounds toward an unrecoverable failure basin (shaded). Recovery is only possible while the trajectory is still outside the point-of-no-return ring (dashed). The AEGIS probe fires the gate within the early window (≤ 30% of trajectory steps, red poi… view at source ↗
Figure 2
Figure 2. Figure 2: Timing doubles recovery at matched compute. Conditional recovered-task rate (RTR) on the weak-policy-failing subset of LIBERO-Spatial (confirmatory n=700; nA-fail=646). At a shared escalation budget (about 38% of steps), AEGIS recovers 10.1% of otherwise-failed trajectories, roughly twice the budget-matched blind (4.6%) and random-trigger (5.1%) controls. The always-strong arm (grey) marks the recovery cei… view at source ↗
Figure 3
Figure 3. Figure 3: AEGIS at run time. The weak policy drives by default (blue path). A logistic-regression probe (the only trained component) reads the weak policy’s frozen layer-15 action-expert activations and emits a per-step risk score. A gate (conformal trigger + early-harm guard + per-episode budget cap) turns that score into a binary escalation decision; on flagged steps, control switches at the next chunk boundary to… view at source ↗
Figure 4
Figure 4. Figure 4: How the gate decides. Stylized per-step view of one trajectory. The probe’s risk score st (indigo) rises as failure approaches. The early-harm guard suppresses any escalation before tmin=0.20 T (hatched); the ≤ 30% band is the probe’s evaluation window where AUROC is read, not a runtime firing cutoff. When st crosses the conformal threshold τ (red), control hands to the stronger policy at the next chunk bo… view at source ↗
Figure 5
Figure 5. Figure 5: Offline probe build and calibration. Everything in this figure happens once, before deployment, and changes nothing at rollout time. A held-out calibration set is rolled out under the frozen weak VLA; a forward hook captures layer-15 action-expert activations ht . The only trained component is a two-layer MLP probe head [720 → 256 → 1] fit to the eventual-failure label on early-window steps. Split-conforma… view at source ↗
Figure 6
Figure 6. Figure 6: Selectivity, per step. Each row is one episode; bar colour encodes the per-step probe risk score st (grey low, red high), and red markers show the steps where the gate fires and control hands to the stronger policy. AEGIS escalates only a small fraction of steps per episode (∼38% in the confirmatory n=700 run) rather than running the stronger policy throughout; the figure also shows a late-detection miss (… view at source ↗
Figure 7
Figure 7. Figure 7: What the controller does, on two real episodes. Each panel is one common￾random-number key (same task, same seed, same initial state) run under two arms. Top rows (Weak, arm A): the deployed weak policy alone; per-step probe risk (st) climbs and the trajectory ends in failure. Bottom rows (AEGIS, arm B): on the identical key, the gate fires (▼) in the early window and control hands to the stronger policy (… view at source ↗
Figure 8
Figure 8. Figure 8: Selectivity, not spend, is the lever. Each arm is placed by relative compute cost (horizontal; weak-policy baseline = 1.0) against recovered-task rate (vertical, confirmatory n=700 data). The budget-matched controls C and D sit in the same cost bracket as AEGIS (B) but far below it, while always-strong reaches near-ceiling recovery only at ≈ 4.6× the compute. At the shared, near-baseline compute budget, AE… view at source ↗
Figure 9
Figure 9. Figure 9: Per-task recovered-task rate, conditioned on the A-failing subset. Confirmatory factorial (n=700 CRN cells): for each of the 10 LIBERO-Spatial tasks, conditional RTR for arms B (targeted), C (budget-matched-blind), and D (random-trigger), with Wilson 95% intervals. The bottom POOLED row shows the across-task pooled diamonds (targeted B near 0.10, the budget￾matched controls C and D near 0.05). All three pr… view at source ↗
Figure 10
Figure 10. Figure 10: Early-window (t ≤ 0.30 T) failure-prediction performance of the hidden-state probe on the Phase-D pilot (leave-one-out out-of-fold; 28 success / 84 failure episodes). Left: ROC of the live action-expert probe, early-window AUROC = 0.738 (the confirmatory run re-estimates this at 0.764 over n=2,792 episodes, clearing the ≥ 0.75 main-run precondition; [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The probe is sharpest at the gate readout. Failure-prediction AUROC of the hidden-state probe as a function of the trajectory fraction the probe is allowed to read, computed on n=2,792 episodes across 10 tasks (every point and band is data, not a schematic). The headline early-window number is read from the weak-policy path before any handoff, so the label (eventual failure under the weak policy) and the … view at source ↗
Figure 12
Figure 12. Figure 12: Conditional recovered-task rate (RTR) by difficulty stratum on the confirmatory factorial [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Replicate-resampling robustness of the primary contrasts. For each (task, seed) cell rolled out on more than one host (212 of 700), the simulator is not bit-identical, so the success bit can differ across draws. Resampling which available single-host-complete draw defines each cell (2,000 iterations) yields the plotted distribution of the three primary RTR gaps. Diamonds are the single-host headline estim… view at source ↗
Figure 14
Figure 14. Figure 14: places the recovery against the always-strong ceiling and the budget-matched controls on a recovery-versus-compute plane [PITH_FULL_IMAGE:figures/full_fig_p024_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Recovered-task rate across arms (Phase-D pilot, exploratory). RTR = Pr[success | arm, A-failing episode], with Wilson 95% intervals on the 56-key pilot. Targeted escalation (B) recovers 65.9% of the pilot episodes the weak policy alone fails; the compute-matched controls, blind escalation (C, 14.6%) and the random-trigger placebo (D, 17.1%), spend the same strong-policy budget yet recover far less (∆RTRB−… view at source ↗
read the original abstract

Long-horizon robot manipulation tends to fail gradually: one bad step degrades the state, and the policy spirals into a basin from which it cannot recover. The failure is often visible before it happens. We introduce AEGIS (Activation-probe Early-warning, Gated Inference Switching), a selective escalation method that uses a lightweight probe on a weak policy's frozen activations to detect high-risk steps while there is still time to act. When the probe flags a step, control switches to a stronger separate policy, but only for the steps that need it. On LIBERO-Spatial, AEGIS recovers 10.1% of the trajectories the weak policy alone loses, versus 4.6% for budget-matched blind escalation and 5.1% for a random-trigger placebo. These gains are significant under one-sided exact paired McNemar tests with Holm-Bonferroni adjustment over three pre-registered contrasts: +5.4pp over blind escalation, p=8.5e-6; +5.0pp over random triggering, p=1.0e-4; paired-trajectory bootstrap CIs exclude zero. AEGIS activates the stronger policy on only 38% of steps, so the lever is timing rather than compute. The probe clears its precondition with an early-window AUROC of 0.764, 95% CI [0.70, 0.84], read from the weak-policy path over the first 30% of trajectory steps before any handoff. We pre-register the full analysis plan, including a conditional recovered-task-rate estimand and explicit kill criteria, and confirm the result on 700 common-random-number episodes per arm, with nA-fail=646.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces AEGIS, a selective escalation method for long-horizon robot manipulation that deploys a lightweight probe on the frozen activations of a weak policy to detect high-risk steps and switch control to a stronger policy only when needed. On LIBERO-Spatial it reports recovering 10.1% of trajectories lost by the weak policy (versus 4.6% for budget-matched blind escalation and 5.1% for a random-trigger placebo), with the differences declared significant under one-sided exact paired McNemar tests (Holm-Bonferroni adjusted) and paired-trajectory bootstrap CIs that exclude zero; the probe activates on 38% of steps and achieves an early-window AUROC of 0.764 on the first 30% of weak-policy trajectory steps.

Significance. If the reported recovery rates and statistical controls hold, the work supplies concrete evidence that timing-based selective escalation can improve reliability of physical AI systems without incurring the full cost of the stronger policy on every step. The pre-registered analysis plan, exact paired tests, explicit placebo and budget-matched controls, and bootstrap CIs are positive features that raise the evidentiary bar for the empirical claim.

major comments (1)
  1. [Methods] Methods section: the manuscript does not supply the training procedure for the activation probe, the weak and strong policies, the data splits, or the raw episode logs; without these details it is impossible to verify the pre-registered separation between probe training and evaluation or the absence of post-hoc exclusions, directly undermining confidence in the reported McNemar p-values and recovered-task-rate estimand.
minor comments (1)
  1. [Abstract] Abstract: the notation 'nA-fail=646' is not defined; it should be expanded to 'number of episodes in which the weak policy fails' for immediate clarity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need for greater methodological transparency. We agree that the current manuscript is insufficiently detailed on these points and will revise accordingly to support verification of the pre-registered analysis.

read point-by-point responses

  1. Referee: [Methods] Methods section: the manuscript does not supply the training procedure for the activation probe, the weak and strong policies, the data splits, or the raw episode logs; without these details it is impossible to verify the pre-registered separation between probe training and evaluation or the absence of post-hoc exclusions, directly undermining confidence in the reported McNemar p-values and recovered-task-rate estimand.

    Authors: We accept the criticism. The revised manuscript will add a dedicated subsection detailing (i) the exact training procedure, loss, optimizer, and hyperparameters for the activation probe, (ii) the training regimes and checkpoints for the weak and strong policies, (iii) the train/validation/test splits with explicit confirmation that probe training data were strictly separated from evaluation episodes, and (iv) the pre-registration identifier together with a statement that the registered analysis plan was followed without post-hoc exclusions. Raw episode logs (state, actions, probe scores, and success flags) will be deposited in a public repository with a DOI and linked in the paper. These additions will allow independent verification of the McNemar tests and the conditional recovered-task-rate estimand. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical benchmark results with no derivations

full rationale

The paper reports direct empirical measurements of recovery rates, AUROC, and statistical tests on the fixed LIBERO-Spatial benchmark using pre-registered analysis, controls, and 700 episodes per arm. No equations, first-principles derivations, fitted parameters renamed as predictions, or self-citation chains are present or load-bearing. The central claims are falsifiable experimental outcomes, not reductions to inputs by construction. This matches the default expectation of no significant circularity for an empirical methods paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical performance measured via pre-registered statistical tests on the LIBERO-Spatial benchmark; the probe's predictive power is taken as given by the reported AUROC.

axioms (1)
  • standard math Standard assumptions underlying the McNemar test and paired bootstrap confidence intervals hold for the trajectory-level comparisons
    Invoked to support the reported p-values and CIs

pith-pipeline@v0.9.1-grok · 5841 in / 1315 out tokens · 41885 ms · 2026-06-28T00:52:01.243988+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 2 canonical work pages

  1. [1]

    Unpacking Failure Modes of Generative Policies: Runtime Monitoring of Consistency and Progress, 2024

    Christopher Agia, Rohan Sinha, Jingyun Yang, Zi-ang Cao, Rika Antonova, Marco Pavone, and Jeannette Bohg. Unpacking Failure Modes of Generative Policies: Runtime Monitoring of Consistency and Progress, 2024. URLhttps://arxiv.org/abs/2410.04640

    arXiv 2024

  2. [2]

    Understanding Intermediate Layers Using Linear Classifier Probes, 2016

    Guillaume Alain and Yoshua Bengio. Understanding Intermediate Layers Using Linear Classifier Probes, 2016. URLhttps://arxiv.org/abs/1610.01644

    Pith/arXiv arXiv 2016

  3. [3]

    Angelopoulos and Stephen Bates

    Anastasios N. Angelopoulos and Stephen Bates. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification, 2021. URLhttps://arxiv.org/abs/2107. 07511

    2021

  4. [4]

    URLhttps://arxiv

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.π0: A Visio...

    Pith/arXiv arXiv 2024

  5. [5]

    RynnVLA-002: A Unified Vision-Language-Action and World Model, 2025

    Jun Cen, Siteng Huang, Yuqian Yuan, Kehan Li, Hangjie Yuan, Chaohui Yu, Yuming Jiang, Jiayan Guo, Xin Li, Hao Luo, Fan Wang, Deli Zhao, and Hao Chen. RynnVLA-002: A Unified Vision-Language-Action and World Model, 2025. URLhttps://arxiv.org/abs/2511.17502

    Pith/arXiv arXiv 2025

  6. [6]

    AURA: Action-Gated Memory for Robot Policies at Constant VRAM, 2026

    Josef Chen. AURA: Action-Gated Memory for Robot Policies at Constant VRAM, 2026. URL https://arxiv.org/abs/2606.02775

    Pith/arXiv arXiv 2026

  7. [7]

    Memory-Bound but Not Bandwidth-Limited: The Physical AI Inference Gap in Batch-1 LLM Decode, 2026

    Josef Chen. Memory-Bound but Not Bandwidth-Limited: The Physical AI Inference Gap in Batch-1 LLM Decode, 2026. URLhttps://arxiv.org/abs/2605.30571

    Pith/arXiv arXiv 2026

  8. [8]

    Lingling Chen, Zongyao Lyu, and William J. Beksi. ReconVLA: An Uncertainty-Guided and Failure-Aware Vision-Language-Action Framework for Robotic Control, 2026. URLhttps: //arxiv.org/abs/2604.16677

    Pith/arXiv arXiv 2026

  9. [9]

    Tibshirani.An Introduction to the Bootstrap

    Bradley Efron and Robert J. Tibshirani.An Introduction to the Bootstrap. Chapman and Hall/CRC, New York, 1993. URLhttps://doi.org/10.1201/9780429246593

    work page doi:10.1201/9780429246593 1993

  10. [10]

    Gaus, Jhon P.F

    Johannes A. Gaus, Jhon P.F. Charaja, and Daniel Haeufle. Confidence-Gated Robot Autonomy: When Does Uncertainty Actually Help?, 2026. URLhttps://arxiv.org/abs/2605.18045

    Pith/arXiv arXiv 2026

  11. [11]

    SAFE: Multitask Failure Detection for Vision-Language-Action Models,

    Qiao Gu, Yuanliang Ju, Shengxiang Sun, Igor Gilitschenski, Haruki Nishimura, Masha Itkina, and Florian Shkurti. SAFE: Multitask Failure Detection for Vision-Language-Action Models,

  12. [12]

    URLhttps://arxiv.org/abs/2506.09937

    arXiv

  13. [13]

    A Simple Sequentially Rejective Multiple Test Procedure.Scandinavian Journal of Statistics, 6(2):65–70, 1979

    Sture Holm. A Simple Sequentially Rejective Multiple Test Procedure.Scandinavian Journal of Statistics, 6(2):65–70, 1979. URLhttps://www.jstor.org/stable/4615733

    arXiv 1979

  14. [14]

    INSIGHT: INference-time Sequence Introspection for Generating Help Triggers in Vision-Language-Action Models, 2025

    Ulas Berk Karli, Ziyao Shangguan, and Tesca Fitzgerald. INSIGHT: INference-time Sequence Introspection for Generating Help Triggers in Vision-Language-Action Models, 2025. URL https://arxiv.org/abs/2510.01389

    Pith/arXiv arXiv 2025

  15. [15]

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, 2025

    Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, 2025. URLhttps://arxiv.org/abs/2502.19645. 28

    Pith/arXiv arXiv 2025

  16. [16]

    Failure-Aware RL: Reliable Offline-to-Online Reinforcement Learning with Self-Recovery for Real-World Manipulation, 2026

    Huanyu Li, Kun Lei, Sheng Zang, Kaizhe Hu, Yongyuan Liang, Bo An, Xiaoli Li, and Huazhe Xu. Failure-Aware RL: Reliable Offline-to-Online Reinforcement Learning with Self-Recovery for Real-World Manipulation, 2026. URLhttps://arxiv.org/abs/2601.07821

    arXiv 2026

  17. [17]

    FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models, 2025

    Zijun Lin, Jiafei Duan, Haoquan Fang, Dieter Fox, Ranjay Krishna, Cheston Tan, and Bihan Wen. FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models, 2025. URLhttps://arxiv.org/abs/2510.01642

    arXiv 2025

  18. [18]

    LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning, 2023

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning, 2023. URLhttps: //arxiv.org/abs/2306.03310

    Pith/arXiv arXiv 2023

  19. [19]

    Note on the Sampling Error of the Difference between Correlated Propor- tions or Percentages.Psychometrika, 12(2):153–157, 1947

    Quinn McNemar. Note on the Sampling Error of the Difference between Correlated Propor- tions or Percentages.Psychometrika, 12(2):153–157, 1947. URLhttps://doi.org/10.1007/ BF02295996

    1947

  20. [20]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots, 2025

    NVIDIA, Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu Lin, Kevin Lin, Guilin Liu, Edith Llontop, Loic Magne, Ajay Mandlekar, Avnish Narayan, Soroush Nasiriany, Scott Reed, You Liang ...

    Pith/arXiv arXiv 2025

  21. [21]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch...

    Pith/arXiv arXiv 2025

  22. [22]

    Schoellig

    Ralf Römer, Adrian Kobras, Luca Worbis, and Angela P. Schoellig. Failure Prediction at Runtime for Generative Robot Policies, 2025. URLhttps://arxiv.org/abs/2510.09459

    arXiv 2025

  23. [23]

    Gordon, and J

    Stéphane Ross, Geoffrey J. Gordon, and J. Andrew Bagnell. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. InProceedings of the 14th International Conference on Artificial Intelligence and Statistics (AISTATS), 2011. URL https://arxiv.org/abs/1011.0686

    Pith/arXiv arXiv 2011

  24. [24]

    SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics, 2025

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, Simon Alibert, Matthieu Cord, Thomas Wolf, and Remi Cadene. SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics, 2025. URLhttps://arxiv.org/abs/2506.01844

    Pith/arXiv arXiv 2025

  25. [25]

    Pre-VLA: Preemptive Runtime Verification for Reliable Vision- Language-Action and World-Model Rollouts, 2026

    Zhen Sun, Yongjian Guo, Haoran Sun, Luqiao Wang, Wei Lu, Jiachi Ji, Shengzhe Ji, Junwu Xiong, and Zhijun Meng. Pre-VLA: Preemptive Runtime Verification for Reliable Vision- Language-Action and World-Model Rollouts, 2026. URL https://arxiv.org/abs/2605. 22446. 29

    2026

  26. [26]

    Todorov, T

    Emanuel Todorov, Tom Erez, and Yuval Tassa. MuJoCo: A Physics Engine for Model-Based Control. In2012 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5026–5033, 2012. URLhttps://doi.org/10.1109/IROS.2012.6386109

    work page doi:10.1109/iros.2012.6386109 2012

  27. [27]

    The Intervention Paradox: Accurate Failure Prediction in Agents Does Not Imply Effective Failure Prevention, 2026

    Rakshith Vasudev, Melisa Russak, Dan Bikel, and Waseem Alshikh. The Intervention Paradox: Accurate Failure Prediction in Agents Does Not Imply Effective Failure Prevention, 2026. URL https://arxiv.org/abs/2602.03338

    arXiv 2026

  28. [28]

    Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies, 2025

    Chen Xu, Tony Khuong Nguyen, Emma Dixon, Christopher Rodriguez, Patrick Miller, Robert Lee, Paarth Shah, Rares Ambrus, Haruki Nishimura, and Masha Itkina. Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies, 2025. URLhttps://arxiv.org/abs/2503.08558

    arXiv 2025

  29. [29]

    FPC-VLA: A Vision-Language- Action Framework with a Supervisor for Failure Prediction and Correction, 2025

    Yifan Yang, Zhixiang Duan, Tianshi Xie, Fuyu Cao, Pinxi Shen, Peili Song, Piaopiao Jin, Guokang Sun, Shaoqing Xu, Yangwei You, and Jingtai Liu. FPC-VLA: A Vision-Language- Action Framework with a Supervisor for Failure Prediction and Correction, 2025. URL https://arxiv.org/abs/2509.04018

    arXiv 2025

  30. [30]

    LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies, 2026

    Yue Yang, Shuo Cheng, Yu Fang, Homanga Bharadhwaj, Mingyu Ding, Gedas Bertasius, and Daniel Szafir. LiLo-VLA: Compositional Long-Horizon Manipulation via Linked Object-Centric Policies, 2026. URLhttps://arxiv.org/abs/2602.21531

    arXiv 2026

  31. [31]

    HELM: Harness-Enhanced Long-horizon Memory for Vision-Language-Action Manipulation, 2026

    Zijian Zeng, Fei Ding, Huiming Yang, and Xianwei Li. HELM: Harness-Enhanced Long-horizon Memory for Vision-Language-Action Manipulation, 2026. URLhttps://arxiv.org/abs/2604. 18791. 30

    2026