arXiv subjects

Juyi Sheng

Publications and source records attributed to Juyi Sheng.

4 recordsLinked to original sources

arXiv2026

CORE: Common Outcome Regularities from Action-Free Visual Demonstrations for Robot Manipulation

Robot imitation learning often relies on costly robot demonstrations, while abundant action-free visual demonstrations, such as human videos, are difficult to use because they lack robot-executable actions and suffer from embodiment gaps. We propose CORE, a policy learning framework that extracts Common Outcome Regularities (CORE) from visual demonstrations. Rather than transferring explicit actions across embodiments, CORE exploits a key observation: although successful trajectories for the same task can be diverse, their terminal states often share stable object configurations, spatial relations, and contact constraints. CORE first trains a terminal outcome encoder with contrastive and auxiliary temporal objectives, then aggregates successful terminal embeddings into visual goal prototypes, and finally injects these prototypes as global goal conditions into robot policies. Compared with language instructions, visual goal prototypes provide more concrete geometric and physical constraints for task completion. Across Meta-World, RoboTwin 2.0, and real-world manipulation, CORE improves the average success rate of the corresponding policy backbones by up to +3.9, +11.1, and +17.0 percentage points, respectively, and outperforms text-conditioned variants under the evaluated settings. The project and code are available at https://logssim.github.io/CORE.github.io/.

cs.RO↗

arXiv2026

Eval-Actions: Fine-Grained Execution Quality Evaluation for Robotic Manipulation

Although Vision--Action (VA) and Vision--Language--Action (VLA) policies have advanced robotic manipulation, their evaluation remains dominated by binary success rates, which obscure process-level differences among executions that complete the same task. We introduce Eval-Actions, a diagnostic evaluation methodology and real-robot benchmark for fine-grained execution-quality assessment of learned manipulation policies. Eval-Actions combines criteria-based Expert Grading (EG), Rank-Guided (RG) labels that align measurable motion indicators with expert rankings, and Chain-of-Thought-style (CoT) annotations that explain observable quality differences. The benchmark contains 13K+ teleoperated and policy-generated real-robot episodes covering 150+ tasks and approximately 52 hours of recordings with RGB-D videos, robot-state trajectories, task descriptions, and success/failure labels. Its densely annotated subset provides EG/RG/CoT supervision for training and evaluation. We further provide AutoEval, a reference multimodal evaluator that predicts quality scores, task outcomes, and diagnostic explanations from RGB temporal evidence and compact kinematic summaries. On the annotated Eval-Actions test split, AutoEval-S achieves Spearman rank correlations (SRCCs) of 0.81 and 0.84 under EG and RG, with success detection accuracies of 90.6% and 91.0%; AutoEval-P reaches 0.70 SRCC under CoT. Analyses of expert consistency, physical-metric baselines, modality ablations, structured generalization, and offline policy ranking show that Eval-Actions provides standardized, interpretable diagnostic signals complementary to success-rate evaluation.

cs.RO↗

arXiv2025

MP1: MeanFlow Tames Policy Learning in 1-step for Robotic Manipulation

In robot manipulation, robot learning has become a prevailing approach. However, generative models within this field face a fundamental trade-off between the slow, iterative sampling of diffusion models and the architectural constraints of faster Flow-based methods, which often rely on explicit consistency losses. To address these limitations, we introduce MP1, which pairs 3D point-cloud inputs with the MeanFlow paradigm to generate action trajectories in one network function evaluation (1-NFE). By directly learning the interval-averaged velocity via the "MeanFlow Identity", our policy avoids any additional consistency constraints. This formulation eliminates numerical ODE-solver errors during inference, yielding more precise trajectories. MP1 further incorporates CFG for improved trajectory controllability while retaining 1-NFE inference without reintroducing structural constraints. Because subtle scene-context variations are critical for robot learning, especially in few-shot learning, we introduce a lightweight Dispersive Loss that repels state embeddings during training, boosting generalization without slowing inference. We validate our method on the Adroit and Meta-World benchmarks, as well as in real-world scenarios. Experimental results show MP1 achieves superior average task success rates, outperforming DP3 by 10.2% and FlowPolicy by 7.3%. Its average inference time is only 6.8 ms-19x faster than DP3 and nearly 2x faster than FlowPolicy. Our project page is available at https://mp1-2254.github.io/, and the code can be accessed at https://github.com/LogSSim/MP1.

cs.RO↗

arXiv2025

GPA-RAM: Grasp-Pretraining Augmented Robotic Attention Mamba for Spatial Task Learning

Fine-grained robotic manipulation often fails when inaccurate initial grasps propagate errors and necessitate complex pose correction. We propose Grasp-Pretraining Augmentation (GPA), which incorporates grasp priors from task demonstrations into imitation policies without additional grasp-pose data or annotation. When added to RVT2, GPA raises the average success rate on RLBench from 79.3% to 84.2%. When added to ACT, it raises success on ALOHA cube transfer and bimanual insertion from 86% and 16% to 98% and 38%, respectively. To offset added computational costs, we develop Robotic Attention Mamba (RAM) for real-time deployment. RAM combines attention-based spatial feature extraction with state-space modeling to capture long-range dependencies efficiently. The resulting GPA-RAM framework supports discrete keyframe prediction and continuous action generation. We evaluate it on four platforms, including physical UR5 and ARX R5 systems. GPA-RAM achieves an average success rate of 87.5% on RLBench, outperforming RVT2 and ARP+ by 8.2 and 2.6 percentage points, respectively. On ALOHA, it achieves 98% success in cube transfer and 56% in bimanual insertion, improvements of 12 and 40 percentage points over ACT, while operating at approximately 71 frames per second. These results demonstrate that GPA-RAM combines precise manipulation with efficient real-time robotic execution. Code is available at https://gpa-ram.github.io/.

cs.RO↗