Predictive Long-Horizon Progress Modeling with Action-Aware Quantification

Tianchi Xu; Yalin Zheng; Yangxin Zhu; Ruonan Ren; Zheng Fu; Yunlong Wang; Yuanxin Zhong

cs.ROOct 23, 2026

Predictive Long-Horizon Progress Modeling with Action-Aware Quantification

Authors: Tianchi Xu, Yalin Zheng, Yangxin Zhu, Ruonan Ren, Zheng Fu, Yunlong Wang, Yuanxin Zhong

Organizations: RIMBOT · Peking University · Tsinghua University

Abstract

Quantitative and dense task progress feedback is essential for robotic manipulation across data curation, policy evaluation, and reinforcement learning. Existing progress annotations often rely on sparse keyframe markings derived from subjective manual judgments, limiting their scalability and cross-task consistency, particularly in long-horizon manipulation. To overcome these challenges, we propose action-aware quantification, a unified progress specification that decomposes long-horizon trajectories into object-level subtasks via a vision-language model (VLM) and assigns weighted action credits to capture fine-grained execution dynamics. Using VLM auto-labeling and human calibration, we construct an 800+ hour training video dataset and LOOM, a long-horizon progress benchmark containing 19,618 human-calibrated evaluation frames. To model these dense progress dynamics, we introduce NimCritic, a predictive progress critic trained via Joint Video-Progress Learning. We further incorporate Historical Progress Perturbation to mitigate sequential drift, and design a contrastive Video-Progress Alignment objective that bridges asymmetric expert streams by mutually grounding intermediate visual dynamics and denoised progress trajectories. NimCritic matches or outperforms state-of-the-art process reward models (PRMs) on the established Video Progress Benchmark (VPB) and LOOM, surpassing GVL (GPT-6-Sol) on VPB by about 65% in terminal success accuracy (TSA). On the large-scale RoboArena policy evaluation benchmark, it also retrieves human-selected successful rollouts with about 97% overall recall. These results establish action-aware quantification as a scalable foundation for dense progress feedback in robotic manipulation.

Figures & tables

Figure 1: NimCritic training architecture: the video expert learns visual dynamics and supplies intermediate features, the progress expert predicts progress and completion, and Video–Progress Alignment connects the intermediate video feature $h$ to reference and predicted progress trajectories using contrastive supervision.

General-Purpose VLMs with GVL
Method	Overall				Single Task				Multi Task
	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$
GVL (Gemini 3.8 Flash)	0.73	0.73	0.73	0.60	0.69	0.69	0.69	0.57	0.74	0.74	0.74	0.61
GVL (GPT-6-Sol)	0.68	0.68	0.68	0.65	0.69	0.69	0.69	0.65	0.68	0.68	0.68	0.65
Specialized Reward Models
ProcVLM 2B	0.48	0.19	0.27	0.26	0.57	0.20	0.29	0.52	0.44	0.19	0.26	0.13

Table 1: Progress estimation on LOOM.

Figure 2: Progress estimation on LOOM: NimCritic follows different rates of advancement across subtasks, with numbered frames marking action boundaries.

General-Purpose VLMs with GVL
Method	Overall		Expert Seen		Expert Unseen		Non-expert Seen		Non-expert Unseen
	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$
GVL (Gemini 3.8 Flash)	0.73	0.47	0.79	0.43	0.77	0.39	0.66	0.54	0.66	0.56
GVL (GPT-6-Sol)	0.73	0.50	0.77	0.43	0.75	0.39	0.71	0.63	0.69	0.58
Specialized Reward Models
ProcVLM 2B	0.00	0.36	0.00	0.27	0.00	0.26	0.15	0.46	0.23	0.48

Table 2: Progress estimation on VPB.

Figure 3: Progress estimation on VPB across hesitation, recovery, expert execution, and failure.

Additional Training Data			VPB (Overall)
AgiBot	Egocentric	Non-Expert	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$
$\times$	$\times$	$\times$	0.62	0.42	0.50	0.42
$\times$	$\checkmark$	$\times$	0.66	0.72	0.69	0.86
$\times$	$\times$	$\checkmark$	0.72	0.56	0.63	0.88
$\checkmark$	$\checkmark$	$\checkmark$	0.73	0.81	0.77	0.82

Table 3: Evaluation of training data compositions on VPB.

Design Components			VPB (Overall)
Joint Video– Progress Learning	Historical Progress Perturbation	Video–Progress Alignment	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$
$\times$	$\times$	$\times$	0.68	0.24	0.35	0.86
$\checkmark$	$\times$	$\times$	0.68	0.53	0.59	0.75
$\checkmark$	$\checkmark$	$\times$	0.74	0.60	0.67	0.87
$\checkmark$	$\checkmark$	$\checkmark$	0.73	0.81	0.77	0.82

Table 4: Ablation study on VPB.

Figure 4: Mechanism analysis on LOOM: the carton-closing example shows (a) task-object attribution, (b) responses to perturbed progress history, and (c) video and progress feature profiles, with shading denoting one sample standard deviation across five denoising runs.

Method	2 policies		4 policies		7 policies		Overall
	R (%) $\uparrow$	HR (%) $\uparrow$	R (%) $\uparrow$	HR (%) $\uparrow$	R (%) $\uparrow$	HR (%) $\uparrow$	R (%) $\uparrow$	HR (%) $\uparrow$
ProcVLM 2B	50.00	52.58	21.43	37.50	20.57	28.86	36.89	45.45
VLAC-CUT	45.27	48.97	42.86	37.50	63.21	69.92	53.03	55.02
LRM	26.78	26.46	7.14	12.50	35.09	34.96	30.16	28.83
Robometer 4B	50.15	51.20	28.57	25.00	62.08	62.60	55.08	54.31
RynnValue 4B	72.49	72.51	21.43	12.50	48.68	56.91	61.56	67.34

Table 5: Policy rollout evaluation on RoboArena.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition
$t$	Video timestamp.
$i,j$	Indices of action occurrences, including revoked and remaining occurrences.
$t_{i}^{\mathrm{s}},t_{i}^{\mathrm{e}}$	Start and end timestamps of action occurrence $i$ , respectively.
$c_{i}$	Physical action class of occurrence $i$ .
$w_{c}$	Fixed weight of physical action class $c$ .
$\alpha_{i}$	Positive credit factor of occurrence $i$ : full, half, or zero credit.

Appendix

Table A1: Notation for the progress specification and model.

Dataset(s)	Sampling weight	Episodes	Hours
Bridge [ 35 ]	20.00%	37,692	70.21
Fractal / RT-1 [ 2 ]	20.00%	86,100	346.81
ARX / DROID / VLABench / LIBERO	16.67%	20,132	93.86
EgoDex [ 16 ]	13.33%	67,043	171.81
SynData (PsiBot)	13.33%	38,059	117.38
AgiBot [ 3 ]	16.67%	4,250	56.87

Appendix

Table A2: Training data composition.

Dataset / collection	Trajectories	Proportion
ARX	151	50.33%
DROID [ 19 ]	31	10.33%
LIBERO [ 23 ]	2	0.67%
AgiBot [ 3 ]	57	19.00%
AgiBotWorld Beta [ 3 ]	22	7.33%
Galaxea [ 15 ]	11	3.67%

Appendix

Table A3: Video-source composition of LOOM.

Segmenter	SegCover-F1 $\uparrow$	Boundary F1 $\uparrow$	ARI $\uparrow$	SegCover-F1 (1:1) $\uparrow$
GPT-6-Sol	0.78	0.35	0.66	0.76
Qwen 3.8 Max	0.79	0.40	0.70	0.77
Gemini 3.8 Flash	0.80	0.41	0.69	0.78

Appendix

Table A4: VLM-based subtask decomposition on LOOM.

General-Purpose VLMs with GVL
Method	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$
GVL (Gemini 3.8 Flash)	0.92	0.92	0.92	0.61
GVL (GPT-6-Sol)	0.96	0.96	0.96	0.65
Specialized Reward Models
ProcVLM 2B	0.87	0.54	0.67	0.13
VLAC-CUT	0.82	0.37	0.51	0.71

Appendix

Table A5: Progress estimation with shared Gemini 3.8 Flash decomposition on LOOM.

Method	Overall				Single Task				Multi Task
	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$	PRC ${}^{+}\!\uparrow$	PRC ${}^{-}\!\uparrow$	PRC-F1 $\uparrow$	TSA $\uparrow$
NimCritic w/o Subtask Decomposition	0.64	0.54	0.59	0.92	0.70	0.64	0.67	0.88	0.61	0.49	0.55	0.93
NimCritic	0.87	0.86	0.86	0.92	0.70	0.64	0.67	0.88	0.95	0.95	0.95	0.93

Appendix

Table A6: Direct and subtask-based scoring with the same NimCritic on LOOM.

Robot Task Progress Estimation

See all 17 papers

ProcVLM: Learning Procedure-Grounded Progress Rewards for Robotic Manipulation
May 9, 2026Youhe Feng, Hansen Shi, Haoyang Li +7Robot Task Progress Estimation Long-Horizon Robotic Manipulation
School of Information, Renmin University of China · Work done during the internship at Zhipu AI. · Zhipu AI +1
ProgressVLA: Progress-Guided Diffusion Policy for Vision-Language Robotic Manipulation
Mar 29, 2026Hongyu Yan, Qiwei Li, Jiaolong Yang +1Robotic Manipulation Vision-Language-Action Models
Peking University · Microsoft Research Asia (MSRA)
ARM: Advantage Reward Modeling for Long-Horizon Manipulation
Apr 3, 2026Yiming Mao, Zixi Yu, Weixin Mao +5Reward Modeling Long-Horizon Robotic Manipulation
LimX Dynamics · Beijing University of Posts and Telecommunications · Zhejiang University

May 9, 2026J/K move · Enter open · S save

ProcVLM: Learning Procedure-Grounded Progress Rewards for Robotic Manipulation

Youhe Feng, Hansen Shi, Haoyang Li, Xinlei Guo+6

School of Information, Renmin University of China · Work done during the internship at Zhipu AI. · Zhipu AI+1

Robot Task Progress Estimation Long-Horizon Robotic Manipulation Multimodal Reward Modeling Robotic Manipulation VLMs for Robotics

Open paper AskPDF arXiv

Long-horizon robotic manipulation requires dense feedback that reflects how a task advances through its procedural stages, not merely whether the final outcome is successful. Existing reward models often rely on trajectory-level success labels or time-based interpolation, which can conflate elapsed time with true task progress and therefore fail to capture unfinished steps, stagnation, and failure states. We present ProcVLM, a progress-aware vision-language model that learns procedure-grounded progress as a dense reward signal for manipulation. Rather than deriving progress from terminal outcomes or temporal proxies, ProcVLM grounds progress estimation in procedural structure and intra-stage visual change, and further adopts a reasoning-before-estimation paradigm that infers the remaining atomic actions before estimating task progress. Specifically, we construct this supervision by synthesizing frame-level subtask-semantic annotations, assigning progress budgets according to subtask structure, and distributing each budget based on intra-subtask visual change. To train ProcVLM at scale, we build a standardized procedural supervision synthesis pipeline and construct ProcCorpus-60M from 30 embodied datasets with 60M annotated frames, from which we derive ProcVQA for procedure-aware pretraining, with progress estimation as the central task alongside action segmentation and future planning. Experiments on ProcVQA and reward-model benchmarks show that ProcVLM improves embodied procedural reasoning and yields more discriminative trajectory-internal progress estimates than representative baselines, supporting its use as a dense reward model for downstream reward-guided policy optimization. Project page: https://procvlm.github.io/

Abstract

Robot Task Progress Estimation

Explore similar work

ProcVLM: Learning Procedure-Grounded Progress Rewards for Robotic Manipulation

ProgressVLA: Progress-Guided Diffusion Policy for Vision-Language Robotic Manipulation

ARM: Advantage Reward Modeling for Long-Horizon Manipulation