acceptodds
Withdrawn from ICLR 2027

RAISE: Learning from Failures by Reweighting Successful Robot Demonstrations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learning from demonstrations is a cornerstone of general-purpose robotic manipulation. Yet success-centered imitation often underuses failure experience, while monotonic progress supervision overlooks local setbacks and retries within successful executions. We introduce RAISE (Risk-Aware Imitation from Successful Experience), a data-centric framework that turns failures into learning priorities rather than imitation targets. RAISE adapts a stage-aware progress model using only trajectories that reach stable completion, with retry supervision capturing local degradation and recovery. The frozen model grounds failure localization in stage- and progress-matched successful experience through temporally corroborated action deviations, changes in success support, and progress regression. Successful retries and localized failure onsets yield risk anchors that guide offline reweighting of existing demonstrations, emphasizing segments with similar pre-action states and positive predicted task advancement. Simulation and real-robot experiments show that RAISE substantially improves overall task success across diverse imitation learners, including conventional behavior cloning, diffusion policies, and vision–language–action models, without changing policy architectures or imitation targets.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robotic manipulation, imitation learning, learning from failures, task progress estimation, failure localization, demonstration reweighting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xAfzOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents RAISE (Risk-Aware Imitation from Successful Experience), a way to use failed robot executions in imitation learning without imitating them. It first adapts a pretrained progress model on successful trajectories, so that it predicts the task stage and how far the task has progressed, and its prediction drops when the robot makes a mistake. The model is then used to find where failed rollouts start to go wrong. These moments, together with the moments just before errors in successful retries, serve as risk points. Similar moments in the existing demonstrations then get higher loss weights, without changing the policy or its action targets. RAISE improves success for several imitation learners in simulation and on a real robot.

Critical strengths
  1. Using failures to decide where to focus, rather than what to imitate, is a sensible idea. Successful demonstrations can contain local errors, and failed rollouts are often correct until a decisive mistake, so simply keeping successes and discarding failures misses useful information. RAISE keeps every demonstration and its action target and only changes how much each sample counts, so it never imitates wrong actions.

  2. It is a plug-in that works across very different learners. Since it only changes offline loss weights, it needs no change to the policy or to deployment. The same procedure improves BC, Diffusion Policy and two π0.5-based VLAs in 37 of 40 task-learner pairs (Table 1), and it also helps on a real robot.

Critical weaknesses
  1. The comparison is not data-matched. The baselines train only on the 300 demonstrations, while RAISE also uses more than ten times as many extra successful trajectories for the progress model, references and calibration, plus several hundred failures (Appendix A.2, StackCube). These extra successes contain actions, so simply adding them to the training set might give a similar gain. This is not tested.

  2. The pipeline is complex, and it is not clear which parts matter most. RAISE has many components and needs task-specific stage labels, but its output is only a weight between 1 and 2 per sample. The ablations do not remove the risk term itself (Table 24), so it is hard to tell how much the failure signal contributes.

Questions for authors
  1. Could you add a baseline that trains on the 300 demonstrations plus the extra clean successes, with uniform weights? If RAISE still clearly wins, the gain would come from the reweighting rather than from the extra data.
  2. In RAISE, a sample gets a higher weight if its state looks like a risky moment and its action moves the task forward. Could you add a variant that ignores risk and weights samples only by how much their action moves the task forward? If this variant does about as well as RAISE, the failures would add little.
  3. The progress model is adapted separately for each task. Would one model adapted jointly on all ten tasks work as well? This would make RAISE cheaper to apply to new tasks. This could be very interesting.
8 comments
AuthorsOpenReview ↗

Response to xAfz (1/4): Training on all auxiliary successes

Thank you for proposing this control. We trained BC and DP with uniform weights on the 300 expert demonstrations plus every auxiliary successful trajectory, using three update budgets. The protocol and results are in Rebuttal Revision, H.2, Table 29.

Within the rebuttal time and compute budget, we use behavior cloning (BC) and Diffusion Policy (DP) on five tasks for detailed data and weight-assignment controls. RAISE weights each learner's per-sample loss, including supervised fine-tuning losses for π0.5 and OpenVLA-7B (Eq. 2; A.4–A.5). Table 24(a) reports BC/DP component controls; Table 24(b) additionally reports π0.5 component ablations on three tasks. Table 1 provides paired π0.5 results, and Table 2 reports the GTP-FA-OpenVLA-7B+RAISE combined system.

Successful-trajectory composition

Pool (complete trajectories unless noted)StackLift
Expert action demonstrations300300
Clean progress-adaptation trajectories1,0001,000
Retry trajectories (retry events)729 (1,000)920 (1,000)
Success-reference trajectories1,0001,000
Calibration trajectories1,0001,000
Auxiliary successful trajectories3,7293,920
All successful action trajectories4,0294,220

Stack denotes StackCube; Lift denotes LiftPegUpright. The four auxiliary success pools are trajectory-disjoint and have compatible action records. Each complete trajectory is counted once, including trajectories with multiple retry events. The separately collected 1,000 clean demonstrations in the 1,300-demonstration study are excluded. Development and final-evaluation trajectories are also disjoint.

Update-budget comparison (Table 29)

The uniform columns report BC or DP trained on all N successful trajectories. The last column reports BC+RAISE or DP+RAISE, respectively, using 300 action demonstrations at U. Values are Stable Success@3 (%), mean±sample SD across training seeds 0/1/2. Each model uses the same 200 evaluation initial conditions per task. Stack and Lift have horizons of 250 and 200 environment steps. Stable success requires the completion criterion at three consecutive evaluation states within an episode.

Task / learnerNUniform UUniform 2UUniform 4U+RAISE U
Stack / BC4,0296.0±1.37.3±1.07.8±1.09.8±1.4
Stack / DP4,02940.3±2.841.0±2.341.5±2.343.8±9.5
Lift / BC4,22066.0±6.167.3±4.368.2±3.773.0±5.9
Lift / DP4,22036.0±4.038.7±3.740.2±3.352.8±5.8

U is 10,000 BC updates or 100,000 DP updates; batch size is 1,024. Within each learner, input and action representations, initialization, normalization, optimizer and execution control are matched. Normalization uses the 300 expert demonstrations.

Expanded-data runs retain optimizer state through 4U; DP also retains EMA state. After U, BC keeps learning rate 0.0003; DP warms up for 500 updates to 0.0001 and then cosine-decays through the extension. Checkpoints at U, 2U and 4U were fixed before evaluation: 12 independent training paths yield 36 checkpoint evaluations and 7,200 episodes.

At U, BC+RAISE and DP+RAISE exceed their uniform counterparts at 4U by 2.0/2.3 percentage points on Stack and 4.8/12.7 on Lift. Differences use unrounded rates. Seed variability is reported in the table, including the 9.5-point SD for Stack DP+RAISE.

The experiment compares two uses of the same auxiliary successes. RAISE additionally uses natural failures to compute offline weights: Stack has 489 natural failures, 352 of which produce confirmed anchors. The methods therefore have different total auxiliary inputs. The fixed-progress-model control in E.5, Table 24(a) isolates the risk factor's contribution. Training and evaluation details are in H.1–H.2.

AuthorsOpenReview ↗

Response to xAfz (2/4): Adding 1,000 clean demonstrations

We also tested a larger action-training dataset with added clean demonstrations on five tasks. With 1,300 action demonstrations, five-task mean stable success is 21.3% for BC and 41.1% for BC+RAISE; DP and DP+RAISE achieve 25.0% and 36.6%. Results are in Rebuttal Revision, H.2, Table 28; seed-level counts and paired gains are in H.4, Tables 33–34.

For each task, we separately collected 1,000 successful, retry-free trajectories using the same demonstrator, state distribution and action interface as the 300 expert demonstrations. The new trajectories are disjoint from those demonstrations and every auxiliary pool. This experiment tests weighting with 1,000 additional retry-free demonstrations. The all-auxiliary-success experiment in H.2, Table 29 tests two uses of the existing auxiliary successes; it uses a different set of added trajectories.

BC and BC+RAISE (Table 28)

TaskBC 300BC+RAISE 300BC 1300BC+RAISE 1300
Stack0.09.82.210.7
Push2.331.86.733.5
Lift61.273.062.774.3
Peg10.343.214.746.7
Apple19.838.020.340.2
Mean18.739.221.341.1

DP and DP+RAISE (Table 28)

TaskDP 300DP+RAISE 300DP 1300DP+RAISE 1300
Stack39.243.841.544.3
Push0.00.03.08.7
Lift24.752.823.853.3
Peg0.210.22.312.3
Apple52.263.354.364.5
Mean23.234.025.036.6

BC and DP use uniform weights; the suffix gives the action-demonstration count. Entries are three-training-seed mean Stable Success@3 (%); Mean weights the five tasks equally. Stack, Push, Lift, Peg and Apple denote StackCube, PushCube, LiftPegUpright, PegInsertionSide and PlaceApple, with horizons 250/50/200/100/100 environment steps. Seeds 0/1/2 each use the same 200 evaluation initial conditions. Stable success requires three consecutive evaluation states satisfying the completion criterion.

BC+RAISE minus BC and DP+RAISE minus DP (Table 34)

Action-demonstration counts are matched within each pair. Gains are in percentage points.

Learner / action NSeed 0Seed 1Seed 2Mean±SD
BC / 30027.015.119.220.4±6.0
BC / 1,30019.619.020.719.8±0.9
DP / 30011.312.88.310.8±2.3
DP / 1,30012.110.812.011.6±0.7

For each seed, we compute within-task differences and average the five tasks equally. Mean±SD summarizes these three seed-level gains using sample SD. The 600 episodes per configuration comprise three models tested on the same 200 initial conditions. Differences use unrounded rates.

Across data scales, policy input representations, architecture, initialization, normalization, update budget and evaluation protocol stay fixed. BC uses 10,000 updates and DP 100,000, with batch size 1,024. Within each task and data scale, BC/BC+RAISE and DP/DP+RAISE share action demonstrations, targets, batch order and execution settings. Push's execution library expands from 300 to 1,300 trajectories, with its neighbor count and fusion rule fixed; each baseline–RAISE pair uses the same library.

Auxiliary progress models, references, calibration sets and risk anchors stay frozen across data scales. RAISE also uses auxiliary success and failure inputs, so matched action-demonstration counts do not imply equal total data budgets.

Adding action data improves the five-task means, with task-level exceptions: Lift DP drops from 24.7% to 23.8%. Push DP and DP+RAISE both achieve 0% with 300 demonstrations; with 1,300 they reach 3.0% and 8.7%. At both data scales, the paired five-task gains of BC+RAISE and DP+RAISE are positive for each training seed. Protocols are in H.1–H.2.

AuthorsOpenReview ↗

Response to xAfz (3/4): Risk, weight assignment and component ablations

We tested the requested control by fixing the progress model and setting the risk factor to 1, with the action-advancement gate and weight schedule unchanged. Rebuttal Revision, E.5, Table 24(a) reports these component results.

BC: 11 configurations

VariantStackPushLiftPegAppleMean
BC (uniform)0.02.361.210.319.818.7
Retry risk only0.00.055.76.314.215.2
Natural-failure risk only4.210.567.534.230.229.3
No retry supervision2.78.358.531.525.725.3
Progress-only localization0.00.058.28.716.516.7
No advancement gate5.726.369.740.234.335.2
Progress-only weighting3.515.564.020.227.226.1
Global shuffle2.212.562.515.221.222.7
Conditional shuffle2.710.761.516.522.522.8
No-retry-risk shuffle2.39.760.714.720.521.6
BC+RAISE9.831.873.043.238.039.2

DP: 11 configurations

VariantStackPushLiftPegAppleMean
DP (uniform)39.20.024.70.252.223.2
Retry risk only36.50.020.50.044.520.3
Natural-failure risk only40.30.047.56.856.330.2
No retry supervision39.80.046.85.342.226.8
Progress-only localization38.20.022.70.048.721.9
No advancement gate39.70.048.58.759.231.2
Progress-only weighting41.20.030.54.756.526.6
Global shuffle38.50.025.72.253.223.9
Conditional shuffle40.50.027.21.250.723.9
No-retry-risk shuffle40.70.023.71.549.223.0
DP+RAISE43.80.052.810.263.334.0

Entries are mean Stable Success@3 (%) across training seeds 0/1/2, with the same 200 evaluation initial conditions per model. Mean weights five tasks equally. StackCube, PushCube, LiftPegUpright, PegInsertionSide and PlaceApple use horizons 250/50/200/100/100 environment steps. Stable success requires completion at three consecutive evaluation states. Each configuration uses 300 action demonstrations, 10,000 BC or 100,000 DP updates, and batch size 1,024.

Control definitions

Progress-only weighting sets the sample-weight risk factor to 1 with the progress model fixed. Progress-only localization retains risk anchors and matching, using progress evidence alone to locate failures.

The two single-risk-source variants retain the full progress model and advancement gate; full RAISE takes the larger of the two risks. No retry supervision sets the retry-loss coefficient to zero, readapts the model, and rebuilds references, calibration, anchors and weights. No advancement gate sets the advancement factor to 1 while preserving anchors and matches.

Global shuffling preserves weight values and multiplicities but permutes their assignment to samples. Conditional shuffling constrains permutations by stage, within-stage progress and advancement. No-retry-risk shuffling exchanges weights only among samples with positive advancement and zero successful-retry risk. All controls retain their learner's action representation, controller, per-sample loss and weight-application schedule.

Paired gains of BC+RAISE or DP+RAISE (Table 34)

Learner / controlGains: seeds 0 / 1 / 2Mean±SD (pp)
BC / progress weighting12.7 / 12.4 / 14.213.1±1.0
BC / global shuffle15.4 / 16.8 / 17.216.5±0.9
DP / progress weighting6.3 / 11.1 / 5.07.5±3.2
DP / global shuffle9.1 / 13.8 / 7.510.1±3.3

For each seed, we subtract the control's success rate from its learner's full RAISE success rate within each task, then average five tasks equally. SD is the sample standard deviation of those three seed-level gains; calculations use unrounded rates. Risk-factor gains are positive for all three five-task seed averages in both learners. For DP, progress-only weighting adds 3.3±1.7 percentage points over uniform training; the risk factor adds a further 7.5±3.2.

BC+RAISE and DP+RAISE have the highest five-task means in their respective tables. Retry-risk-only and progress-only localization have lower means than uniform training. All DP variants achieve 0% on Push at 50 steps. These mechanism results cover BC and DP on five tasks. Definitions, seed counts and paired statistics are in E.5, Table 24(a), and H.4, Tables 32–34; the common protocol is in H.1.

Reviewer_duwTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RAISE, a method that uses failure experience to reweight successful demonstrations for robot imitation learning. It first adapts a progress model using successful trajectories, including retry segments, to capture task advancement and local setbacks. It then analyzes failed trajectories against successful references to identify potentially risky situations. Action chunks from successful demonstrations receive higher training weights when they begin in similar situations and are predicted to advance the task. The method preserves the policy architecture and action targets, with all auxiliary processing performed offline and no additional model calls during deployment. Experiments across 10 simulation tasks and five real-world tasks, together with component ablations, demonstrate improvements across multiple policy families.

Critical strengths
  1. A practical way to use failure experience. RAISE connects failure experience with successful demonstrations to identify where the policy needs more focused training, while retaining successful actions as supervision. The method can be integrated into existing policies without changing their architecture or adding model calls during deployment, making it applicable across different policy families.

  2. Broad empirical support. The authors evaluate the method across multiple simulation tasks, policy families, and real-world tasks, with improvements across a broad range of settings. The component ablations further examine the contributions of the main design choices. Together, these experiments provide convincing evidence that the method’s benefits extend beyond a single task or policy.

Critical weaknesses
  1. Limited comparison with closely related weighting methods. The main controlled comparisons are against uniform training. Without direct comparisons to ReTVL or simpler progress-based weighting under matched conditions, it remains unclear how much additional benefit the failure localization and risk matching mechanisms provide. The component ablations are useful but do not fully establish the advantage over these alternatives.

  2. Unclear benefits relative to the additional resources required. Although policy action supervision uses 300 demonstrations, RAISE also requires substantial auxiliary trajectories, task-specific annotations, and progress-model adaptation. The paper does not sufficiently quantify these costs or establish how policy improvements change as auxiliary resources decrease. Where suitable action labels are available, comparing against directly incorporating the additional successful trajectories into policy training would help determine whether the proposed pipeline is a worthwhile use of these resources.

Questions for authors

See the weaknesses.

9 comments
AuthorsOpenReview ↗

Response to duwT (1/4): Weighting and data-curation comparisons

Thank you for requesting ReTVL and progress-weighting comparisons and a quantitative analysis of auxiliary resources. The Rebuttal Revision includes the detailed protocols and results below.

Within the rebuttal time and compute budget, we use behavior cloning (BC) and Diffusion Policy (DP) on five tasks for detailed data and weight-assignment controls. RAISE weights each learner's per-sample loss, including supervised fine-tuning losses for π0.5 and OpenVLA-7B (Eq. 2; A.4–A.5). Table 24(a) reports BC/DP component controls; Table 24(b) additionally reports π0.5 component ablations on three tasks. Table 1 provides paired π0.5 results, and Table 2 reports the GTP-FA-OpenVLA-7B+RAISE combined system.

1. Five-task comparison (H.1, H.3, Table 30)

Stack, Push, Lift, Peg and Apple denote StackCube, PushCube, LiftPegUpright, PegInsertionSide and PlaceApple, with horizons of 250/50/200/100/100 environment steps. Entries are mean Stable Success@3 (%) over three independently trained seeds; each model uses the same 200 initial conditions. Stable Success@3 requires completion at three consecutive evaluation states. Mean weights tasks equally.

BC results:

MethodStackPushLiftPegAppleMean
BC (uniform)0.02.361.210.319.818.7
BC+AWR (progress)5.231.213.540.338.525.7
BC+DemInf6.330.022.345.037.228.2
BC+Demo-SCORE7.740.370.024.731.234.8
BC+ReTVL (retry-boundary)6.231.223.023.20.016.7
BC+RAISE9.831.873.043.238.039.2

DP results:

MethodStackPushLiftPegAppleMean
DP (uniform)39.20.024.70.252.223.2
DP+AWR (progress)40.00.029.56.760.327.3
DP+DemInf41.20.029.56.261.227.6
DP+Demo-SCORE42.31.247.59.253.530.7
DP+ReTVL (retry-boundary)41.50.029.88.248.525.6
DP+RAISE43.80.052.810.263.334.0

BC+RAISE and DP+RAISE have the highest five-task means. Task-level leaders differ: BC+Demo-SCORE leads on Push, BC+DemInf on Peg and BC+AWR on Apple. On Push, DP+Demo-SCORE reaches 1.2%, compared with 0.0% for DP+RAISE at 50 steps.

2. Method-specific inputs and weights (H.3, Table 31)

Within each learner, methods share the policy configuration, update budget and evaluation protocol. Each starts from 300 candidate expert demonstrations; curation changes action-data membership, and auxiliary inputs differ by method. Counts below are shared across BC, DP and policy seeds. DS = Demo-SCORE; E/O = retained expert/added online-success trajectories; T/R = value-training trajectories/retry events.

TaskDemInf EDS EDS Oγ×10⁴ReTVL T/R
Stack225150141315729/1000
Push2252421104144600/18
Lift2253002082637920/1000
Peg2252991424079669/1000
Apple2252981986196922/995
  • AWR: uses RAISE's adapted, frozen progress model. Endpoint progress differences ΔV estimate advantages: z=(ΔV−median(ΔV))/(IQR(ΔV)+1e-5), w=min(20,exp(z)). Invalid scores receive weight 1. Temperature 1 and cap 20 are fixed. On Lift, BC uses unit weights for the first half of training, then introduces AWR weights linearly.
  • DemInf: trains state/action β-VAEs on nonterminal pairs from the 300 expert trajectories, applies KSG scoring with k=5,6,7, and retains 75% (225 trajectories). Trajectory identifiers break ties. Push's execution-neighbor library uses the retained trajectories.
  • Demo-SCORE: an independently trained, unweighted BC policy collects 100 non-test, 250-step trajectories at each of four checkpoints (2500/5000/7500/10000 updates). The first three groups each train a candidate classifier; the fourth selects the classifier and epoch by validation cross-entropy. Groups are trajectory-disjoint. γ is the selected classifier's mean training-trajectory score. Retained experts must exceed γ; added online trajectories must also succeed. BC and DP share the 400 collected trajectories and selection.
  • ReTVL: independently adapts Robometer-4B for 500 updates/task. The retry-boundary adaptation retains value-increment weighting and uses physical retry starts or geometric-criterion decreases as correction-start events. Terminal success/failure supplies value supervision. Push uses 300 expert plus 300 online trajectories. Lift uses 884 ultimately successful and 36 ultimately failed trajectories; RAISE's separate Lift retry set has 920 ultimately successful trajectories. Each Lift set contains 1000 retry events.

Total data usage differs by method. H.3 gives the weighting formulas; E.5, Table 24(a) isolates the risk factor with the progress model fixed.

AuthorsOpenReview ↗

Response to duwT (2/4): Risk controls and seed variation

1. Risk and weight-assignment controls

The Rebuttal Revision reports these controls in E.5, Table 24(a), and H.4, Tables 32–34. Table 24(a) and Table 32 use 300 action demonstrations and 200 common initial conditions/model. Table 24(a) reports three-seed mean Stable Success@3 (%), requiring completion at three consecutive evaluation states. Stack/Push/Lift/Peg/Apple use 250/50/200/100/100 steps. Cells list BC/DP results under each training rule; Avg. weights tasks equally.

All 11 configurations from Table 24(a):

ControlStackPushLiftPegAppleAvg.
BC/DP (uniform)0.0/39.22.3/0.061.2/24.710.3/0.219.8/52.218.7/23.2
Retry risk only0.0/36.50.0/0.055.7/20.56.3/0.014.2/44.515.2/20.3
Natural-failure risk only4.2/40.310.5/0.067.5/47.534.2/6.830.2/56.329.3/30.2
No retry supervision2.7/39.88.3/0.058.5/46.831.5/5.325.7/42.225.3/26.8
Progress-only localization0.0/38.20.0/0.058.2/22.78.7/0.016.5/48.716.7/21.9
No advancement gate5.7/39.726.3/0.069.7/48.540.2/8.734.3/59.235.2/31.2
Progress-only weighting3.5/41.215.5/0.064.0/30.520.2/4.727.2/56.526.1/26.6
Global shuffle2.2/38.512.5/0.062.5/25.715.2/2.221.2/53.222.7/23.9
Conditional shuffle2.7/40.510.7/0.061.5/27.216.5/1.222.5/50.722.8/23.9
No-retry-region shuffle2.3/40.79.7/0.060.7/23.714.7/1.520.5/49.221.6/23.0
BC+RAISE/DP+RAISE9.8/43.831.8/0.073.0/52.843.2/10.238.0/63.339.2/34.0

Progress-only weighting fixes the progress model, advancement gate and weight schedule, and sets the risk factor to 1. Progress-only localization still builds risk anchors but locates failures using progress evidence alone. Global shuffle preserves all weight values and multiplicities while permuting sample assignments. Conditional shuffle restricts permutations by stage, within-stage progress and advancement; no-retry-region shuffle acts on positive-advancement samples with zero retry risk.

Retry-only and natural-failure-only variants retain the full progress model and advancement gate. Removing retry supervision sets its loss coefficient to zero, refits the progress model and rebuilds downstream auxiliary quantities. Removing the advancement gate sets its factor to 1 with anchors and matches fixed.

2. Paired gains and seed variation (Table 34)

Gains are BC+RAISE minus the listed BC control, or DP+RAISE minus the listed DP control. For each seed, we average within-task differences over five tasks, then report mean±sample SD across three seeds. Values are percentage points, calculated before rounding. BC/DP (300 or 1300) denote unweighted training with matched action-data size. Demo-SCORE uses its selected expert and online-success data.

Control policySeed 0/1/2 gainsMean±SD
BC (300)27.0/15.1/19.220.4±6.0
BC (progress weighting)12.7/12.4/14.213.1±1.0
BC (global shuffle)15.4/16.8/17.216.5±0.9
BC+Demo-SCORE4.3/3.5/5.44.4±1.0
BC (1300)19.6/19.0/20.719.8±0.9
DP (300)11.3/12.8/8.310.8±2.3
DP (progress weighting)6.3/11.1/5.07.5±3.2
DP (global shuffle)9.1/13.8/7.510.1±3.3
DP+Demo-SCORE1.6/7.9/0.43.3±4.0
DP (1300)12.1/10.8/12.011.6±0.7

For DP, progress-only weighting adds 3.3±1.7 points over unweighted DP; the risk factor adds 7.5±3.2, giving a total paired mean gain of 10.8 points. The five-task risk gain is positive for every seed in both BC and DP.

Table 32 reports task-level mean±sample SD with 300 demonstrations:

TaskBCBC+RAISEDPDP+RAISE
Stack0.00±0.009.83±1.4439.17±3.0643.83±9.46
Push2.33±2.3131.83±1.260.00±0.000.00±0.00
Lift61.17±30.1773.00±5.8924.67±6.1152.83±5.80
Peg10.33±2.3643.17±5.390.17±0.2910.17±2.47
Apple19.83±4.1938.00±4.9252.17±5.0363.33±4.73

RAISE reduces seed variation for Lift BC but increases it for Stack DP. Both Push DP configurations remain at 0% at 50 steps. The variance effect is task-dependent.

3. Onset localization (C.4, Table 17(b))

With the final progress checkpoint, detections and logged decision indices fixed, C2 refines the onset using action evidence within a five-step look-back.

OnsetRecallFPRNLEAcc@3Acc@5Acc@10
Unrefined69.013.00.4431029.237.842.6
C2 refined69.013.00.4134036.441.749.4

Acc@k is localization within k steps, with all 1000 controlled failures in the denominator. Recall uses those failures; FPR uses 1000 clean successes. Rates are percentages. NLE averages onset error divided by 50, assigning 50-step error to misses; 50 is the normalization scale, not the execution horizon. These are offline localization metrics, distinct from policy success.

AuthorsOpenReview ↗

Response to duwT (3/4): Direct action training with extra successes

1. Use the same auxiliary successes for action supervision

Rebuttal Revision H.2, Table 29 compares direct action supervision with auxiliary use of the same successful trajectories. Stack combines 300 action demonstrations with 3729 auxiliary successes (1000 clean, 729 retry-containing, 1000 reference and 1000 calibration), giving 4029 trajectories. Lift has 3920 auxiliary successes, including 920 retry-containing trajectories, giving 4220. Each retry set contains 1000 events; events add no trajectories. All action records are used. Development and final evaluation are trajectory-disjoint.

Unweighted BC and DP train continuously to 4U, with U/2U/4U checkpoints fixed before evaluation. U is 10000 BC updates or 100000 DP updates. Optimizer state and DP's EMA state are retained. After U, BC keeps learning rate 3e-4; DP warms up for 500 updates to 1e-4, then cosine-decays over the extension. BC+RAISE and DP+RAISE use 300 action demonstrations at U, with auxiliary successes and natural failures for offline weighting.

Table 29 reports mean±sample SD of Stable Success@3 (%), three seeds and 200 common initial conditions/model. Stack uses 250 steps and Lift 200. Stable success requires completion at three consecutive evaluation states. N is the unweighted policy's action-trajectory count; the last column is BC+RAISE or DP+RAISE, matching the row.

Task/policyNUniform UUniform 2UUniform 4URAISE U
Stack/BC40296.0±1.37.3±1.07.8±1.09.8±1.4
Stack/DP402940.3±2.841.0±2.341.5±2.343.8±9.5
Lift/BC422066.0±6.167.3±4.368.2±3.773.0±5.9
Lift/DP422036.0±4.038.7±3.740.2±3.352.8±5.8

At U, BC+RAISE/DP+RAISE exceed unweighted BC/DP at 4U by 2.0/2.3 points on Stack and 4.8/12.7 on Lift, calculated before rounding. RAISE also uses natural-failure evidence, so total auxiliary inputs differ. E.5, Table 24(a) isolates the risk factor with the progress model fixed.

The expanded-data experiment has 12 independent training paths (2 tasks×2 learners×3 seeds), each evaluated at three checkpoints: 36 checkpoint evaluations and 7200 episodes.

2. Add 1000 separately collected clean demonstrations per task

These trajectories use the same demonstrator, state distribution and action interface as the 300 demonstrations, and are disjoint from them and all auxiliary pools. Each BC–BC+RAISE or DP–DP+RAISE pair shares initialization, architecture, action targets, normalization, update budget and execution settings. Push's execution-neighbor library expands with action-data size and is shared within each pair. Auxiliary models, references, calibration and risk anchors stay fixed across sizes.

Complete Table 28: three-seed mean Stable Success@3 (%) at Stack/Push/Lift/Peg/Apple horizons of 250/50/200/100/100. Mean weights tasks equally.

BC results:

Policy (action N)StackPushLiftPegAppleMean
BC (300)0.02.361.210.319.818.7
BC+RAISE (300)9.831.873.043.238.039.2
BC (1300)2.26.762.714.720.321.3
BC+RAISE (1300)10.733.574.346.740.241.1

DP results:

Policy (action N)StackPushLiftPegAppleMean
DP (300)39.20.024.70.252.223.2
DP+RAISE (300)43.80.052.810.263.334.0
DP (1300)41.53.023.82.354.325.0
DP+RAISE (1300)44.38.753.312.364.536.6

Table 33's four 1300-demonstration configurations: stable-success counts for seeds 0/1/2, each out of 200. BC and DP denote unweighted training.

PolicyStackPushLiftPegApple
BC1/8/412/15/13124/127/12529/33/2643/41/38
BC+RAISE21/17/2667/69/65144/149/15393/97/9080/82/79
DP85/83/817/6/544/48/515/8/1109/109/108
DP+RAISE89/89/8818/17/17105/107/10825/25/24134/124/129

With 1300 demonstrations, BC+RAISE and DP+RAISE have paired five-task gains of 19.8±0.9 and 11.6±0.7 points over BC and DP, respectively (Table 34). Extra demonstrations increase the overall means, with task-specific exceptions: unweighted DP on Lift decreases from 24.7% to 23.8%. BC+RAISE and DP+RAISE have higher five-task means at both data sizes.

The separately collected 1000 demonstrations are excluded from the all-auxiliary experiment above. The two experiments test distinct data controls; repeated 300-demonstration results reuse the same models. Full protocols and results are in H.2, Tables 28–29, and H.4, Tables 33–34.

Reviewer_nRneOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RAISE, which uses failed robot attempts to decide which parts of the successful demonstrations to emphasize during training.

A progress model is trained on successful demonstrations only. It predicts the task stage and a progress score. A retry loss makes predicted progress drop during a mistake and rise after recovery. The frozen model finds where each failed rollout went wrong by comparing each step with similar moments from successful runs. A failure is flagged when the actions look unusual or the scene departs from what successful runs show. These failure points and the pre-mistake states from successful retries are used as anchors. Each sample in the training demonstrations gets a weight. The weight is higher when the sample is close to an anchor in the same stage and when progress increases over its action chunk. In simulation, average success rises by about 10 points over standard training for BC, Diffusion Policy, π0.5 and GTP-FA-π0.5 on ten ManiSkill3 tasks. On five real-robot tasks, success rises from 11.2% to 26.4% for π0.5 and from 76.8% to 84.0% for GTP-FA-π0.5.

Critical strengths

Practical and easy to adopt. RAISE uses failed rollouts, which are usually discarded, without the risk of imitating failed actions. It changes only the loss weights, leaves the policy architecture, training data and action targets unchanged, adds no cost at test time, and does not need online RL. This matters because the method can be added to almost any existing imitation pipeline.

Broad evaluation. The evaluation covers four policy types (BC, Diffusion Policy, π0.5 and GTP-FA-π0.5), ten simulation tasks and five real-robot tasks. Success improves in 37 of 40 simulation settings and on all five real-robot tasks. The paper also includes ablations of each component and separate evaluations of the progress model and the failure detector. This matters because it shows the gains are not limited to one policy or task.

Critical weaknesses

Missing baselines and an unequal data budget. Uniform weighting is the only controlled baseline; the comparisons in Table 2 reuse published numbers under different data and protocols. There is no matched comparison to other weighting or data-curation methods such as AWR-style weighting, DemInf or Demo-SCORE. Table 24 ablates components inside RAISE, but two simple controls are missing: weighting by progress gain alone (no risk term), and the same weights randomly shuffled across samples. In addition, RAISE uses much more data than the baselines. For StackCube (App. A.2), the baselines train on 300 demonstrations, while RAISE additionally uses 1000 clean successes and 1000 retry events from 729 successful trajectories to train the progress model, a 1000-trajectory success reference bank and a 1000-trajectory calibration set for failure detection, and 489 natural failed rollouts to build the failure anchors. In total this is thousands of additional successful trajectories (about 3,700 if the pools are disjoint), versus 300 for the baselines. This is consequential because, without these comparisons, the paper does not show that the gains come from the failure information, which is its main claim.

The source of the gains is unclear. For π0.5 on StackCube, the mean weight is 1.11 and about half the samples keep a weight of exactly 1 (Table 22). For BC, task advancement is measured over one step (Ha=1H_a = 1, Eq. 15) instead of ten, so its weights should be even closer to 1, yet BC gains as much as the other policies (+11.4 points vs. +10.1 to +11.2). The paper also reports no per-seed results or standard deviations for policy success. The real-robot baseline numbers are reused from Xu et al. (2026) rather than rerun, and with 50 trials per task several real-robot gains (e.g., 92→94 on Orange-to-Tray, 86→90 on PokeCube) are within trial-to-trial variation. Finally, the weighting thresholds (δz\delta_z, δu\delta_u, δV\delta_V, β\beta) are fixed across all tasks; the paper states they were fixed before policy training but not how they were chosen, and there is no sensitivity analysis. This is consequential because it is hard to tell whether the improvements are larger than the variation between training runs, and what in the method causes them.

Questions for authors
  1. How does a baseline trained on the training demonstrations plus the extra successful trajectories compare with RAISE?

  2. How does RAISE compare with other weighting or data-curation methods, such as AWR-style weighting, DemInf or Demo-SCORE?

  3. How does RAISE compare with weighting by progress gain alone (with no risk term), and with the same weights randomly shuffled across samples?

  4. What is the weight distribution for each policy, especially BC?

  5. The paper does not report the extra data used for the real-robot experiments. How many successful trajectories, retries and failed rollouts were used, and how were the failures collected?

11 comments
AuthorsOpenReview ↗

Response to nRne (1/5): Direct training on the auxiliary successes

Thank you for requesting this data-budget control. The Rebuttal Revision reports it in H.2, Tables 28–29.

Within the rebuttal time and compute budget, we use behavior cloning (BC) and Diffusion Policy (DP) on five tasks for detailed data and weight-assignment controls. RAISE weights each learner's per-sample loss, including supervised fine-tuning losses for π0.5 and OpenVLA-7B (Eq. 2; A.4–A.5). Table 24(a) reports BC/DP component controls; Table 24(b) additionally reports π0.5 component ablations on three tasks. Table 1 provides paired π0.5 results, and Table 2 reports the GTP-FA-OpenVLA-7B+RAISE combined system.

Stable Success@3 requires completion at three consecutive evaluation states. Success rates below are percentages averaged over seeds 0/1/2; each model uses the same 200 evaluation initial conditions. SD is the sample standard deviation across trained models.

Auxiliary-success pools (H.2; revised Table 29). The disjoint pools contain these trajectory counts:

PoolStackCubeLiftPegUpright
Clean auxiliary successes10001000
Retry-containing successes729920
Success references10001000
Success calibration10001000
Auxiliary total37293920
Including 300 action demonstrations40294220

Each retry pool has 1000 events; events are counted separately from trajectories. BC and DP train on the action records from all 4029/4220 trajectories. BC+RAISE and DP+RAISE use 300 action demonstrations, weighted using auxiliary evidence.

Direct action supervision from all auxiliary successes (revised Table 29). Stack/Lift horizons are 250/200 steps. U is 10,000 BC or 100,000 DP updates. Base denotes BC or DP; +RAISE denotes BC+RAISE or DP+RAISE, respectively.

Task/learnerBase UBase 2UBase 4U+RAISE U
Stack BC6.0±1.37.3±1.07.8±1.09.8±1.4
Stack DP40.3±2.841.0±2.341.5±2.343.8±9.5
Lift BC66.0±6.167.3±4.368.2±3.773.0±5.9
Lift DP36.0±4.038.7±3.740.2±3.352.8±5.8

U/2U/4U checkpoints are predefined on continuous training paths, retaining optimizer and DP exponential-moving-average states. After U, BC keeps learning rate 0.0003; DP warms up over 500 updates to 0.0001 and cosine-decays through the extension. There are 12 training paths, 36 checkpoint evaluations and 7200 episodes.

At U, BC+RAISE/DP+RAISE exceed their respective 4U baselines by 2.0/2.3 points on Stack and 4.8/12.7 on Lift, computed before rounding. This compares two uses of the same auxiliary successes. RAISE also uses natural failures: 489 Stack trajectories, 352 producing confirmed risk anchors. E.5, Table 24(a) isolates the risk factor with the progress model fixed.

Separately collected action demonstrations (revised Table 28). An additional 1000 retry-free successes per task are disjoint from the 300 demonstrations and all auxiliary pools; they do not enter the 4029/4220 comparison. Each cell reports BC/DP or BC+RAISE/DP+RAISE; suffixes give action-trajectory counts.

TaskBase300+RAISE300Base1300+RAISE1300
Stack0.0/39.29.8/43.82.2/41.510.7/44.3
Push2.3/0.031.8/0.06.7/3.033.5/8.7
Lift61.2/24.773.0/52.862.7/23.874.3/53.3
Peg10.3/0.243.2/10.214.7/2.346.7/12.3
Apple19.8/52.238.0/63.320.3/54.340.2/64.5
Mean18.7/23.239.2/34.021.3/25.041.1/36.6

Push/Peg/Apple denote PushCube/PegInsertionSide/PlaceApple, with horizons 50/100/100; Stack/Lift retain 250/200. Means weight tasks equally. Both data scales share initialization, update budgets, normalization and evaluation settings. Auxiliary models, references, calibration and anchors stay frozen. Push's execution library expands with action data and is shared by each matched BC or DP pair.

With 1300 trajectories, BC+RAISE exceeds BC by 19.8±0.9 points and DP+RAISE exceeds DP by 11.6±0.7 (Table 34). Differences are computed within each task and seed, then averaged over tasks; SD describes the three seed-level gains.

Task horizon H versus 250 steps (revised Table 27). With 300 demonstrations, the same checkpoints and trajectory prefixes give the following mean Stable Success@3 (%). Main Table 1 uses 250 steps; the controlled comparisons above use H.

Task/learnerHBase H/250+RAISE H/250
Stack BC2500.0/0.09.8/9.8
Push BC502.3/3.331.8/34.5
Lift BC20061.2/63.873.0/74.7
Peg BC10010.3/15.543.2/49.3
Apple BC10019.8/23.538.0/43.3
Stack DP25039.2/39.243.8/43.8
Push DP500.0/0.70.0/4.0
Lift DP20024.7/28.052.8/56.0
Peg DP1000.2/0.710.2/14.2
Apple DP10052.2/57.363.3/69.7

The horizon changes; the trained models remain fixed.

AuthorsOpenReview ↗

Response to nRne (2/5): Weighting and data-curation baselines

The Rebuttal Revision compares AWR, DemInf, Demo-SCORE and a retry-boundary adaptation of ReTVL in H.3, Tables 30–31.

BC denotes behavior cloning; DP denotes Diffusion Policy. Success-rate entries are three-seed mean Stable Success@3 (%), requiring completion at three consecutive evaluation states. Models trained with seeds 0/1/2 use the same 200 evaluation initial conditions. Stack/Push/Lift/Peg/Apple denote StackCube/PushCube/LiftPegUpright/PegInsertionSide/PlaceApple; horizons are 250/50/200/100/100 steps. Each method starts from 300 candidate demonstrations. Policy configurations, update budgets and evaluation settings are fixed; curation methods select or add action trajectories.

BC results (revised Table 30).

PolicyStackPushLiftPegAppleMean
BC0.02.361.210.319.818.7
BC+AWR5.231.213.540.338.525.7
BC+DemInf6.330.022.345.037.228.2
BC+Demo-SCORE7.740.370.024.731.234.8
BC+ReTVL6.231.223.023.20.016.7
BC+RAISE9.831.873.043.238.039.2

DP results (revised Table 30).

PolicyStackPushLiftPegAppleMean
DP39.20.024.70.252.223.2
DP+AWR40.00.029.56.760.327.3
DP+DemInf41.20.029.56.261.227.6
DP+Demo-SCORE42.31.247.59.253.530.7
DP+ReTVL41.50.029.88.248.525.6
DP+RAISE43.80.052.810.263.334.0

Means weight tasks equally. BC+RAISE and DP+RAISE have the highest five-task means. BC+Demo-SCORE and DP+Demo-SCORE lead on Push; BC+DemInf leads on Peg and BC+AWR on Apple. Against BC+Demo-SCORE and DP+Demo-SCORE, BC+RAISE and DP+RAISE gain 4.4±1.0 and 3.3±4.0 points, respectively (Table 34): mean±sample SD of three seed-level five-task gains, calculated before rounding. SD describes training-seed variation, not a confidence interval.

Method-specific inputs (revised Table 31). Demo-SCORE collects 400 non-test trajectories per task. “Expert + online” counts retained expert and added successful online trajectories; γ is the selection threshold. ReTVL counts are value-model training trajectories/retry events, separate from action-training counts.

TaskDemo-SCORE expert + onlineγ×10⁴ReTVL traj./events
Stack150 + 141315729/1000
Push242 + 1104144600/18
Lift300 + 2082637920/1000
Peg299 + 1424079669/1000
Apple298 + 1986196922/995

Implementation and data use (H.3).

AWR shares RAISE's frozen, adapted Robometer-4B progress model and derives exponential weights from endpoint progress differences. Differences are centered by their median and divided by the interquartile range plus 0.00001; temperature is 1 and weights are capped at 20. Invalid scores receive weight 1. On Lift, BC+AWR uses unit weights in the first half of training, then linearly introduces AWR weights.

DemInf trains state/action β-VAEs on the 300 candidate demonstrations and uses KSG scores with neighbor counts 5/6/7. Fixed 75% retention keeps 225 trajectories per task; ties are resolved by trajectory ID. Push's execution library uses this retained set.

Demo-SCORE's unweighted BC collector contributes 100 trajectories at each of four checkpoints. The first three groups train candidate classifiers; the fourth selects the classifier and epoch by validation cross-entropy. Training and validation trajectories are disjoint. γ is the selected classifier's mean training-trajectory score. Experts must exceed γ; added online trajectories must also succeed. BC, DP and all their seeds share the collected pool and selection.

ReTVL independently adapts Robometer-4B for 500 updates per task. It retains value-increment weighting, uses physical retry starts or geometric-criterion decreases as correction boundaries, and uses terminal outcomes for supervision. Push uses 300 expert plus 300 online trajectories. Lift's 920 value-training trajectories contain 884 successes and 36 failures; RAISE's separate 920-trajectory retry set contains only successes.

These comparisons fix policy and evaluation settings while retaining each method's data requirements. The published-system entries in Table 2 use their respective evaluation protocols.

AuthorsOpenReview ↗

Response to nRne (3/5): Risk controls and per-seed success counts

The Rebuttal Revision reports risk controls and raw success counts by training seed in E.5, Table 24(a), and H.4, Tables 32–34.

BC denotes behavior cloning; DP denotes Diffusion Policy. Stable Success@3 requires completion at three consecutive evaluation states. All models use 300 action demonstrations. Models are trained independently with seeds 0/1/2; each model uses the same 200 evaluation initial conditions. The 600 episodes comprise three models on 200 conditions. Stack/Push/Lift/Peg/Apple denote StackCube/PushCube/LiftPegUpright/PegInsertionSide/PlaceApple; horizons are 250/50/200/100/100 steps.

Fixed-progress-model controls (revised Table 24(a)). Each cell reports BC/DP mean success (%); full RAISE denotes BC+RAISE/DP+RAISE. Means weight tasks equally.

RuleStackPushLiftPegAppleMean
Unit weights0.0/39.22.3/0.061.2/24.710.3/0.219.8/52.218.7/23.2
PW3.5/41.215.5/0.064.0/30.520.2/4.727.2/56.526.1/26.6
GS2.2/38.512.5/0.062.5/25.715.2/2.221.2/53.222.7/23.9
CS2.7/40.510.7/0.061.5/27.216.5/1.222.5/50.722.8/23.9
ZS2.3/40.79.7/0.060.7/23.714.7/1.520.5/49.221.6/23.0
Full RAISE9.8/43.831.8/0.073.0/52.843.2/10.238.0/63.339.2/34.0

PW uses progress-only weighting: it sets the risk factor to 1 and fixes the progress model, advancement gate and weight schedule. GS globally permutes weight–sample assignments, preserving weight values and multiplicities. CS restricts permutation by stage, within-stage progress and advancement. ZS permutes only advancing samples with zero retry risk. Within each learner, all controls share the action representation, controller, loss and weight-application schedule.

PW isolates the risk increment with a fixed progress model; GS/CS/ZS test weight assignment. BC+RAISE and DP+RAISE have the highest five-task means among these controls. DP Push remains at zero for all six rules at 50 steps.

BC success counts for seeds 0/1/2 (revised Table 33). Each individual count is out of 200.

Policy/controlStackPushLiftPegApple
BC0/0/02/10/253/163/15119/17/2644/45/30
BC+RAISE18/18/2364/61/66133/149/15694/74/9179/84/65
BC+PW8/7/629/33/31133/128/12337/40/4454/54/55
BC+RAISE (GS)8/4/125/21/29128/122/12531/30/3042/41/44

DP success counts for seeds 0/1/2 (revised Table 33). Each individual count is out of 200.

Policy/controlStackPushLiftPegApple
DP85/77/730/0/036/60/521/0/095/115/103
DP+RAISE96/101/660/0/0100/119/9818/26/17116/134/130
DP+PW85/80/820/0/057/65/619/11/8116/113/110
DP+RAISE (GS)73/81/770/0/056/47/514/6/3106/108/105

For each task and seed, BC and BC+RAISE are trained independently from identical initial parameters, as are DP and DP+RAISE. Batch order is matched within each pair. The counts show task- and seed-specific variation: on Stack at seed 2, DP+RAISE has fewer successes than DP despite its higher three-seed mean.

Paired five-task gains (revised Table 34). BC columns compare BC+RAISE with BC using the listed control rule; DP columns compare DP+RAISE with the corresponding DP control. Within each seed, compute task-wise differences and then average over five tasks. Entries are percentage points; SD is the sample standard deviation of the three seed-level gains.

Control ruleBC seeds 0/1/2BC mean±SDDP seeds 0/1/2DP mean±SD
Unit weights27.0/15.1/19.220.4±6.011.3/12.8/8.310.8±2.3
PW12.7/12.4/14.213.1±1.06.3/11.1/5.07.5±3.2
GS15.4/16.8/17.216.5±0.99.1/13.8/7.510.1±3.3

For DP, progress-only weighting adds 3.3±1.7 points over unweighted DP; the risk factor adds 7.5±3.2 points over DP+PW. Their mean increments sum to 10.8 points. Risk increments are positive for the three seed-level five-task averages; individual task–seed outcomes vary. All differences are calculated before rounding. The reported SDs quantify training-seed variation, not confidence intervals.

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

Related papers

Discussion (0)

Sign in to comment.