Changing the holding reference during a fixed-duration wait can reverse the ranking of visual recovery methods. We introduce AuditEmbodied, a paired execution and trajectory evaluation framework that independently varies holding references and waiting duration while keeping recovery executors frozen. A 960-execution factorial on 20 initialization seeds compares CAD-feedback and landmark-based recovery from identical post-wait physical states within each holding and waiting condition. Protected completion requires successful recovery under the declared execution constraints. At a 1 s wait with registered wrist references, their protected completion counts change from 6/40 versus 19/40 under registered holding to 34/40 versus 22/40 under position-only holding. Trajectory analysis localizes the main separation to grasp formation: small observation-time target residuals precede large tracking errors. Joint diagnostics and controlled interventions characterize persistent tracking constraints. On fresh seeds, gradual position reanchoring improves CAD recovery's long-wait completion by 42.5 percentage points and reduces mean complete recovery time by 6.06 s relative to registered holding. Additional evaluations with pretrained policies, handover, and a mass-matched geometry characterize holding-reference sensitivity across executors and physical regimes. These findings establish holding reference as an explicit evaluation condition for comparing visual recovery performance.
Figures & tables
Figure 1: Waiting targets reverse the recovery-method ranking. (a) Registered and position-only holding start from a shared fault prefix and track different position references. Within each holding condition, frozen CAD-feedback and landmark executors share the same post-wait physical state. (b) At a fixed 1 s wait with unchanged wrist references, their protected-completion ranking reverses. Each condition contains 40 fault executions from 20 paired seeds.
Executor
Hold target
Protected completion Y500
Healthy
Complete time (s)
S
L
S
L
CAD feedback
Registered
36/40
6/40
40/40
27.01
35.58
CAD feedback
Position only
34/40
34/40
40/40
26.58
27.97
CAD feedback
Rotation only
35/40
9/40
40/40
27.03
34.41
CAD feedback
Position + rotation
36/40
36/40
40/40
27.00
28.21
Landmark
Registered
21/40
19/40
38/40
30.74
31.67
Table 1: Shared-prefix recovery comparison. S/L : 0.05/1.00 s initial waiting; 40 faults per cell from 20 paired seeds. Healthy pools 20 controls per wait. Complete time includes all fault paths, failures, visual services, and retries.
Figure 2: Error decomposition and recovery stages. Each 1 s, registered-wrist condition retains all 40 faults. Signed projections sum to closure error; black ticks show its mean norm. Stage counts include later retries.
Figure 3: Independent waiting-dose validation. Twenty seeds yield 800 executions; error bars resample whole seeds, retaining both fault sides. Complete time includes failures, services, retries, and verification. All endpoint healthy controls protect.
Policy
Boundary
Registered holding
Position-only holding
Interaction [95% CI]
S
L
S
L
Diffusion Policy
32
18/20
19/20
17/20
19/20
+5.0 [0.0,15.0]
Diffusion Policy
272
19/20
19/20
19/20
19/20
0.0 [0.0,0.0]
LSTM–GMM
32
9/20
11/20
13/20
13/20
−10.0 [-40.0,25.0]
LSTM–GMM
272
8/20
8/20
7/20
8/20
+5.0 [-15.0,25.0]
Table 2: Holding interventions on released policies. Native Transport completion C uses 20 seeds per cell. S/L : 0.05/1.00 s extra physical waiting. Interactions are within-policy, within-boundary position-by-wait contrasts in percentage points; 95% intervals resample whole seeds.
Operating regime
Long-wait outcome
Interaction (pp)
Interpretation
Registered
Position only
Core CAD feedback
6/40
34/40
+75.0
Large reference-dependent response
Arrival-adaptive CAD
8/40
37/40
+67.5
Persists with bounded tracking
Released CNN grasp recovery
17/20
18/20
+5.0
Capable recovery; small response
Mass-matched geometry, CAD
37/40
36/40
0.0
Dual grasps formed; balance-limited completion
Handover, local recovery
36/40
36/40
0.0
Better tracking; unchanged completion
Table 3: Scope of holding-reference sensitivity. Interactions are within-cohort position-by-wait contrasts in percentage points. Released CNN uses native recovery R ; other rows use Y500 . Core/adaptive CAD report the registered-wrist slice; Eq. 4 averages across wrist settings.
Long-horizon manipulation requires robots to retain task-relevant history while revising conclusions no longer supported by physical evidence. Retaining observations alone does not specify which completed conditions should be withdrawn after a disturbance or which obligations require repair. We introduce REMM, Revisable Embodied Memory for Manipulation, a supervisory memory layer for a frozen G0.5 vision-language-action executor. REMM separates an append-only evidence history from revisable task state and explicitly links derived relations and completed conditions to their supporting facts. When accepted evidence changes a supporting fact, it withdraws dependent conclusions and schedules targeted recovery at verified skill boundaries without updating the executor. Across 51 physical trials, including operator-assisted runs, REMM improves mean recorded task progress over the strongest baseline from 57% to 91% in intervention recovery and from 43% to 70% in sequence recall. Complementary video diagnostics show improved grounding of retained history into the next instruction, including comparisons with π0.5-compatible high-level planner adaptations. These results support revisable memory as a practical interface for continued manipulation without policy retraining. Project: https://revanthsenthil.github.io/remm/.
Reliable robotic manipulation requires timely intervention to correct emerging deviations and restore progress after execution errors. However, recovery methods based on repeated vision-language reasoning or iterative online optimization can incur substantial latency, delaying intervention. To address these challenges, we introduce RESETTLE(Robotic rEcovery through diSagrEement-Triggered reTrievaL and Efficient Corrective Control), a model-agnostic framework that provides computationally efficient recovery at the action-execution interface of frozen robot policies. RESETTLE triggers recovery when two action proposals independently sampled under identical conditioning persistently disagree. It retrieves a same-task demonstration reference using an adapted V-JEPA encoder and combines a state-servo prior with a guarded visual residual to execute one corrective action without online trajectory optimization or additional vision-language reasoning, then returns control to the base policy. Across six base policies in simulation, RESETTLE achieves up to 8.70%, 6.28%, and 6.83% absolute success-rate gains on LIBERO-Plus, Meta-World, and RoboCasa Tabletop, respectively, with further improvements on four real-world tasks using two policies. In QwenPI-based comparisons, its monitoring-and-recovery computation latency is 74.04%--93.57% lower than VoLoAgent's monitoring-and-planning latency for grasp and place tool calls. It also raises Harness VLA's LIBERO-Pro Swap success from 42% to 50%, demonstrating compatibility with high-level agentic planning. Code available at: https://github.com/JIA-Lab-research/RESETTLE
Yuxin Chen, Senqiao Yang, Zixuan Wang +6
The Hong Kong University of Science and Technology · The Chinese University of Hong Kong · SmartMore +2
Simulation enables scalable training of Vision-Language-Action policies by using privileged experts to generate visual demonstrations without requiring every trajectory to be collected through manual teleoperation. However, such pipelines typically retain successful demonstrations while failed rollouts are discarded, even though they expose precisely the off-nominal states from which recovery must be learned. We introduce Kintsugi-VLA, a framework for converting failed rollouts into targeted synthetic recovery data by exploiting exact state restoration and branching in simulation. For a fixed privileged expert, we define interventional recoverability as the probability of completing the original task after the simulator is restored to a given state, estimate it using adaptive Monte Carlo continuations with pointwise Wilson intervals, and characterize its non-monotonic evolution along failed trajectories. These estimates identify an observed terminal low-recoverability frontier-the point after which measured recoverability remains below a threshold-which is then used to select informative recovery starting states. In a simulated Franka manipulation task, targeted recovery data yield aggregate SmolVLA recovery success of 34.6% and 38.4% under difficulty- and frame-budget matching, respectively, 5.8 and 6.7 percentage points above uniform sampling within the same recovery window. The same ordering is observed under disturbed end-to-end execution and shifted clutter and physics conditions, while clean-task success decreases from 76.8% to 74.7%. Kintsugi-VLA demonstrates how failed simulator rollouts can be transformed from discarded experience into structured recovery-training data through direct interventional measurement.
Ivan Snegirev, Elizaveta Semenyakina, Dmitrii Maliukov +2
Intelligent Space Robotics Laboratory, Skolkovo Institute of Science and Technology