Robotics · 2026-10-26

Predictive Long-Horizon Progress Modeling with Action-Aware Quantification

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

arXiv:2610.32776PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors propose action-aware quantification for dense, absolute long-horizon progress feedback in robotic manipulation by assigning weighted action credits within object-level subtasks and handling temporary attempts and explicit credit revocation. They train NimCritic, a dual-diffusion-transformer predictive progress critic via Joint Video–Progress Learning with Historical Progress Perturbation and Video–Progress Alignment. It matches or outperforms prior process reward models on VPB/LOOM and retrieves successful rollouts on RoboArena.

Why it matters

Long-horizon process reward models require consistent progress magnitude (not just ordering) and must handle non-monotonic execution, including progress regression and revocation of earlier credit. Action-aware quantification provides a unified absolute progress specification with dense supervision from VLM pre-annotation plus human verification, and NimCritic ties predictive video modeling to both progress and completion outputs.

Method

  • Action-aware quantification: define object-level subtasks, locate temporal breakpoints with action boundaries, assign fixed class weights and weighted action credit; forward effects get full credit, engaged-only attempts get temporary half credit, and later revocations remove referenced earlier credit, then compose retained local credit across subtasks.
  • Train NimCritic: a dual-diffusion-transformer world reward model with separately supervised progress and completion outputs; use Joint Video–Progress Learning with Historical Progress Perturbation and Video–Progress Alignment to connect intermediate video features to annotated progress trajectories.
  • Evaluation uses learned progress rank correlations (forward and time-reversed) and terminal success accuracy (TSA), and includes LOOM (19,618 frames) and VPB split metrics plus RoboArena rollout retrieval recall/hit rate.
Abstract (from arXiv)

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.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-23 · Indexed by Pub-AI: 2026-10-26 UTC

preprint

Predictive Long-Horizon Progress Modeling with Action-Aware Quantification

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

Preprint posted: 2026-10-23 · Original record ↗

arxiv
2610.32776
Explicit source versions (1)
  • v1 · 2026-10-23
  1. Preprint available

    Source date: 2026-10-23 · First detected: 2026-10-27 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

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.