Chronos

A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation

Yulin Zhou, Yimeng Wang, Nengyu Wang, Shaojia Xing, Shiyun Tu, Xiang Li, Jingkai Zhang, Ningbo Jiang, Yuankai Lin, Hua Yang, Xiangrui Zeng, Zhouping Yin

Huazhong University of Science and Technology

Currently under review at IEEE Transactions on Robotics (T-RO)

Cover Blocks: Hidden Color-Location Memory

From different starting block orders, the robot must uncover the blocks in red-green-yellow order. pi0.5 confuses similar observations and stalls, while Chronos remembers the hidden color positions and finishes the task.

pi0.5: Single-Frame Aliasing
Chronos: Layout 1
Chronos: Layout 2
Chronos: Layout 3
73.6%
RMBench average success
+62.4
points over pi0.5
10×
fewer parameters than pi0.5
78%
real-world average success

Overview

Chronos addresses non-Markovian long-horizon manipulation by treating observation history as the latent state of the policy dynamics. At each physical control step, Chronos forms one state-representative token from observation and proprioception, propagates the full causal history with a selective state space model, generates a multimodal coarse action prior with IMLE, and refines that prior with a second-order Schrodinger-inspired acceleration bridge.

Chronos Framework

1

Full-History State Encoding

One token is aligned with each physical control step. A selective state space model propagates the complete causal history as the policy state.

2

IMLE Action Prior

A history-conditioned IMLE generator captures multimodal coarse action chunks, keeping multiple feasible modes available for downstream refinement.

3

Second-Order Action Bridge

A Schrodinger-inspired acceleration field refines the coarse action prior into smoother and more physically grounded robot motion.

Chronos framework figure

Full-history state propagation, multimodal action-prior generation, and acceleration-driven second-order refinement.

Key Results

Memory-Dependent Control

On RMBench, Chronos reaches 73.6% average success, outperforming pi0.5 by +62.4 absolute percentage points with 10× fewer parameters.

Compact Yet Strong

Chronos surpasses the memory-aware VLA Mem-0 by 22.8 points while using over 30× fewer parameters.

Real-World Validation

In real-world dual-arm experiments using a single RGB camera, Chronos achieves 78% average success and 72% on memory-dependent tasks.

Real-World Deployment

Real-world experiments compare pi0.5 and Chronos side by side. Chronos preserves hidden task phase and history-dependent information, while pi0.5 often skips intermediate stages or prematurely executes the final action.

Task 1: Long-Horizon Sequence

pi0.5
Chronos

Task 2: Visible Manipulation

pi0.5
Chronos

Task 3: Memory-Dependent Extension

pi0.5
Chronos

BibTeX

@article{zhou2026chronos,
  title={Chronos: A Physics-Informed Full-History Framework for Non-Markovian Long-Horizon Manipulation},
  author={Zhou, Yulin and Wang, Yimeng and Wang, Nengyu and Xing, Shaojia and Tu, Shiyun and Li, Xiang and Zhang, Jingkai and Jiang, Ningbo and Lin, Yuankai and Yang, Hua and Zeng, Xiangrui and Yin, Zhouping},
  journal={arXiv preprint arXiv:2606.30318},
  year={2026}
}