Predictive Long-Horizon Progress Modeling with Action-Aware Quantification

1 Introduction

• Existing PRMs use relative comparisons, manually staged interpolation, or VLM-generated subtask budgets, leaving cross-task magnitude and non-monotonic long-horizon execution insufficiently specified.Relative comparisons order trajectories without specifying cross-task progress magnitude, while elapsed-time interpolation varies with execution speed.Source: §1 · p. 2
• Action-aware quantification defines object-level subtasks and assigns weighted action credit, including temporary credit for engaged attempts and revocation when later actions undo earlier achievements.VLM pre-annotation and human verification produce structured action records for rule-based dense progress computation.Source: §1 · p. 2
• The training dataset contains 253,276 episodes spanning 800+ hours, while LOOM covers 300 cross-embodiment trajectories and 19,618 human-calibrated evaluation frames.LOOM covers differing advancement rates, hesitation, regression, recovery, and incomplete execution.Source: §1 · p. 2
• NimCritic combines predictive video learning with separately supervised progress and completion outputs, using Historical Progress Perturbation and Video–Progress Alignment to support sequential estimates and cross-modal correspondence.Joint Video–Progress Learning ties observed-state evolution to advancement toward a subtask goal.Source: §1 · p. 2
• NimCritic reaches 0.77 overall PRC-F1 and 0.82 TSA on VPB, surpassing GVL (GPT-6-Sol) on TSA by about 65%; it leads external baselines on LOOM and retrieves RoboArena successes with 96.72% overall recall.On LOOM, it leads external baselines across all four metrics in the overall and multi-task groups, with 0.86 and 0.95 PRC-F1, respectively; RoboArena hit rate is 98.09%.Source: §1 · p. 2

Robotic Process Reward Models

• Robotic PRMs use VLM-based rewards, relative progress, temporal distance, or action values, but these distinct signals leave a unified absolute progress scale across tasks unspecified.NimCritic defines quantitative absolute progress through observable action effects within object-level subtasks.Source: §2 · p. 3

Human-Aligned Reward Learning

• Human-aligned reward methods include trajectory comparisons, qualitative stages, procedural budgets, directional judgments, and recovery segmentation, yet remain limited in quantifying non-monotonic long-horizon behavior.Action-aware quantification formalizes full credit, temporary attempt credit, and revocation, with VLM pre-annotation and human verification.Source: §2 · p. 3

World Models

• World models supply temporal priors through future-observation or latent-state prediction, while explicit links between intermediate video representations and quantitative progress remain a distinct modeling opportunity.NimCritic introduces contrastive Video–Progress Alignment and Historical Progress Perturbation to connect predictive video learning with reward estimation.Source: §2 · p. 3

3.1 Architecture

• NimCritic pairs a video expert that learns visual dynamics with a progress expert that predicts progress and completion, then composes subtask predictions into global progress.The critic predicts within-subtask progress and completion from a subtask goal and observed video window; Figure 2 shows contrastive alignment connecting an intermediate video feature to reference and predicted progress trajectories.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2 Progress Annotation Pipeline

• Action-aware annotation assigns action-class weights and task-effect-dependent credits, including temporary half credit for engaged-only attempts and revocation of referenced earlier credit.For incomplete subtasks, the VLM identifies necessary remaining or restorative actions to set a fixed budget; human verification refines boundaries and credit timing before deterministic post-processing computes dense values.Source: §3.2 · pp. 4–5

3.3 VLM-Based Subtask Decomposition

• A VLM partitions long-horizon videos into ordered, contiguous object-level subtasks, preserving slips and retries on the same object within one segment while separating distinct object operations.Segments are paired with local instructions and scored in parallel; duration-weighted composition retains earlier subtasks’ exit contributions while later subtasks reserve duration weight.Source: §3.3 · p. 5

3.4 Predictive Critic

• The predictive critic jointly models video and progress latents, conditions predictions on the subtask instruction, observed video, and historical progress anchor, and reuses encoded video states during inference.Historical Progress Perturbation replaces training anchors within bounded limits and recomputes relative targets from unchanged absolute annotations; Video–Progress Alignment uses contrastive supervision to connect observed-video features with annotated and predicted progress curves.Source: §3.4 · pp. 5–6

4.1 Experimental Setup

• Evaluations use each baseline’s native inference pipeline and observation interface; NimCritic uses Gemini 3.8 Flash for subtask decomposition.The reported metrics include forward and reverse progress rank correlation, their harmonic combination, and terminal success accuracy.Source: §4.1 · p. 7

4.2 Long-Horizon Progress Estimation on LOOM

• On LOOM, NimCritic has the highest overall scores across all four metrics: 0.87 PRC+, 0.86 PRC−, 0.86 PRC-F1, and 0.92 TSA.On multi-task trajectories, it reaches 0.95 PRC+, 0.95 PRC−, 0.95 PRC-F1, and 0.93 TSA.Source: Table 1 · p. 7, §4.2 · p. 7, Figure 3 · p. 8

4.3 Progress Estimation on the Public VPB Benchmark

• 0.77 overall PRC-F1 and 0.82 TSA make NimCritic the top specialized reward model across all four VPB splits.It exceeds GVL (GPT-6-Sol) in TSA, 0.82 vs. 0.50; unseen-task TSA is 0.84 for expert and 0.78 for non-expert rollouts.Source: Table 2 · p. 8, §4.3 · p. 9

4.4 Effect of Training Data Composition

• Adding egocentric human video raises PRC− from 0.42 to 0.72, PRC-F1 from 0.50 to 0.69, and TSA from 0.42 to 0.86.Adding non-expert robot trajectories raises PRC+ from 0.62 to 0.72, PRC-F1 to 0.63, and TSA to 0.88; the full mixture reaches 0.77 PRC-F1 and 0.82 TSA.Source: Table 3 · p. 9, §4.4 · p. 9

4.5 Ablation Study

• JVPL raises PRC-F1 from 0.35 to 0.59; adding HPP yields 0.67 PRC-F1 and 0.87 TSA, while adding VPA yields 0.77 PRC-F1 and 0.82 TSA.In the carton-closing analysis, target-object attribution rises from 6.43% for the frozen backbone to 23.32% for NimCritic; video and progress features have Pearson r = 0.93.Source: Table 4 · p. 10, §4.5 · p. 10, Figure 5 · p. 10

4.6 Successful-Rollout Retrieval for Policy Evaluation

• 96.72% overall recall and 98.09% hit rate on RoboArena exceed RoboDopamine-GRM-2 8B’s 90.66% recall and 93.66% HR.Recall measures recovered human-selected successful rollouts; hit rate measures positive sessions with at least one recovered success.Source: §4.6 · pp. 10–11, Table 5 · p. 11

• Object-level subtasks and weighted action credits assign numerical progress across attempts, regressions, and recoveries.The specification supports VLM pre-annotation with human verification and the human-calibrated LOOM benchmark.Source: §5 · p. 11
• NimCritic combines VLM subtask decomposition with joint video-progress learning, historical progress perturbation, and video-progress alignment.The conclusion says these components ground progress predictions in predictive visual dynamics.Source: §5 · p. 11
• The framework supports long-horizon progress estimation and retrieval of successful rollouts for policy evaluation.The paper frames the contributions as a scalable foundation for dense progress feedback across robotic manipulation applications.Source: §5 · p. 11

• Table A1 summarizes the mathematical symbols and operators used in the progress specification and model.Source: Table A1 · p. 15, §A.1 · p. 15

• The six-source training dataset contains 253,276 episodes and 856.95 hours of manipulation video.The table reports sampling weights, episode counts, and raw video hours by source.Source: Table A2 · p. 16, §A.2 · p. 16
• LOOM contains 300 trajectories: 171 single-task and 129 multi-task, with 6,487 and 13,131 human-calibrated evaluation frames, respectively.The imported block comprises 184 reannotated videos from ARX, DROID, and LIBERO.Source: Table A3 · p. 16, §A.2 · p. 16

A.3 Model, Training, and Inference Details

• Training windows use nine RGB frames at 2 Hz and five progress–completion latent positions, with 128,000 updates and an effective batch size of 192.The visual stream starts from Wan2.2-TI2V-5B; task and subtask instructions use a frozen UMT5-XXL encoder.Source: §A.3 · p. 16
• At inference, five observed frames produce two video latent frames; overlapping windows denoise five-step trajectories for 20 Euler steps and aggregate predictions with inverse-distance weights.Windows advance by two positions, and later windows condition autoregressively on the running distance-weighted average of prior predictions.Source: §A.3 · p. 16
• GVL prompts Gemini 3.8 Flash and GPT-6-Sol to predict progress for up to 30 shuffled frames, then restores timeline order.Its per-image predictions are reused for forward and reverse evaluations; qualitative analyses also describe attention attribution, target-object ROI attribution, endpoint-relative features, and history sensitivity.Source: §A.3 · p. 17

B Evaluation Metrics

• Per-trajectory Spearman rank correlations compare the reference progress curve with forward and reverse predictions aligned to the original timeline.Average ranks are used for ties, and each directional correlation is averaged over its finite support.Source: §B.1 · p. 17
• PRC-F1 is the signed harmonic combination of corpus-level forward and reverse correlations, each aggregated with trajectory-duration weights.The same duration weights aggregate terminal success accuracy (TSA).Source: §B.1 · p. 17
• TSA compares predicted completion decisions with reference binary completion labels defined by the unified progress threshold.The passage specifies duration-weighted aggregation for TSA.Source: §B.1 · p. 17

B.2 Policy Rollout Evaluation

• Across 3,872 RoboArena sessions and 10,750 rollouts, positive-rollout/positive-session counts are 676/582, 14/8, and 530/246 for two-, four-, and seven-policy groups, and 1,220/836 overall.Trajectory-level recall uses human-selected successful rollouts as the reference; session-level hit rate is defined over positive sessions.Source: §B.2 · p. 17

B.3 Subtask Decomposition Quality

• SegCover-F1 is the harmonic mean of directional, segment-length-weighted coverage scores between reference and predicted subtask partitions.The partitions comprise nonempty half-open frame intervals, and interval IoU is the intersection-over-union of their frame sets.Source: §B.3 · pp. 17–18
• The one-to-one SegCover variant uses an optimal Hungarian assignment maximizing matched segment IoU, with unmatched segments contributing zero to the numerator.Unmatched segments remain in the denominator.Source: §B.3 · p. 18
• Boundary F1 measures matched reference and predicted internal boundaries within an adaptive tolerance, while ARI measures frame-level cluster agreement corrected for chance.Boundary pairs are greedily matched in order of increasing temporal distance; ARI uses the frame-pair Rand index and its size-preserving permutation expectation.Source: §B.3 · p. 18

C Additional Experimental Results

• Gemini 3.8 Flash achieves 0.80 SegCover-F1, 0.78 one-to-one SegCover-F1, and 0.41 Boundary F1; Qwen 3.8 Max achieves 0.70 ARI.Gemini is selected as the common decomposition frontend, with cached segments used for NimCritic and shared-segment critic comparisons.Source: Table A4 · p. 18, C Additional Experimental Results · p. 18
• 0.95 vs. 0.86 PRC-F1 and 0.93 vs. 0.71 TSA: under shared decomposition, NimCritic leads specialized reward models; its PRC-F1 is comparable to GPT-6-Sol at 0.95 vs. 0.96, with higher TSA at 0.93 vs. 0.65.The comparison uses cached Gemini 3.8 Flash segments and duration-weighted global composition on 129 multi-task LOOM episodes.Source: C Additional Experimental Results · p. 18, Table A5 · p. 19
• 0.55 to 0.95 PRC-F1: subtask decomposition lifts NimCritic’s multi-task score compared with direct whole-video scoring.Both variants have identical single-task predictions; subtask conditioning and duration-weighted composition provide object-level subgoals and reset each subtask anchor to zero.Source: Table A6 · p. 19, C Additional Experimental Results · p. 19