APM: two paths meeting at an anchor

APM

ANCHORED PATH MATCHING

Distilling World Action Models
for Efficient Bimanual Manipulation
via Anchored Path Matching

Jialong Liu1,2Junjie He2Yandu Chen1,3Bin Zhou1,4Heng Su1,5Peijun Tang1Jianan Wang1Changhao Chen2
1Astribot2The Hong Kong University of Science and Technology (Guangzhou) 3Harbin Institute of Technology4The Hong Kong University of Science and Technology5Chongqing University
Corresponding author: changhaochen@hkust-gz.edu.cn

WATCH THE OVERVIEW

APM: method and experiments.

4:15 · English narration & captions

Method animations, RoboTwin results, ablations, video prediction, and real-robot demonstrations. English narration and on-screen captions guide the overview.

RoboTwin footage shows a selected place_can_basket Random case with matched physical initial states (V/A = 1/1; video/action CFG = 5/1). The comparison uses a Stage I EMA4000 continued-distillation candidate and an exact-case LingBot-VA rerun; this candidate differs from the budget-matched Continued ablation. This qualitative example was selected for APM success and candidate failure, rather than to estimate aggregate success rates.

FEWER STEPS. PAIRED SUPERVISION.

Learn from where the student actually goes.

Coarse denoising can move a student away from its teacher’s trajectory. Anchored Path Matching (APM) restarts the teacher at a student-reached state, pairing a completed-output target with instantaneous field matching at that same state and time. One checkpoint supports 1, 2, and 4 steps, with no teacher at deployment.

Code · Models

Eight-step teacher versus naive four-step generation and four-step APM in bimanual manipulation
80.9%One-step RoboTwin average success
91.4%Four-step RoboTwin average success
21.3×Real-robot speedup vs. eight-step teacher

Abstract

World Action Models jointly generate future video and robot actions through iterative denoising. Few-step distillation reduces the number of model evaluations, but its coarse updates can move the student away from the teacher’s trajectory. We propose Anchored Path Matching (APM), which supervises endpoint predictions and instantaneous direction fields at states reached by the student. From each state, we continue the teacher to zero noise for the restart endpoint and match the teacher and student directions at the same detached state and time. Modality-routed adapters learn both objectives from the shared starting state. When the two gradients conflict, anchor-priority updates preserve the endpoint gradient and remove the opposing component of the field gradient within each modality. One checkpoint supports 1-, 2-, and 4-step joint generation. On RoboTwin 2.0, APM improves single-step success over continued distillation by 2.0 percentage points on Clean and 1.4 points on Random. On two real-robot tasks, it improves success over the one-step teacher and reduces control-cycle latency by 21.3× relative to the eight-step teacher.

One teacher continuation. Two complementary targets.

APM first learns a finite-interval video–action transition map through teacher-integrated distillation. It then freezes the base and trains modality-routed adapters to correct predictions from states encountered along the student’s own trajectories.

APM pipeline: restart teacher from student-reached states, learn endpoint and instantaneous-field targets with anchor-priority gradients, and deploy at multiple sampling budgets
APM pairs endpoint and local-transition targets from the same teacher continuation. Anchor-priority correction coordinates their gradients; deployment uses only the student. Click any figure to view it at full resolution.
01 / REACH

Start on the student’s path

Collect a state from the student’s generation trajectory and hold it fixed for correction.

02 / ANCHOR

Pair field and final targets

Continue the frozen teacher to zero noise for the endpoint, then match its instantaneous direction at the shared state and time.

03 / COORDINATE

Prioritize the endpoint

Remove the opposing component of the local gradient when it conflicts with the endpoint gradient, within each modality.

One checkpoint, three sampling budgets.

APM is trained on Clean data only and evaluated directly on RoboTwin 2.0’s Clean and domain-randomized Random settings. Across 1-, 2-, and 4-step inference, it improves average success over continued base distillation. Select a matched video/action budget below.

RoboTwin 2.0 success rates by matched video/action sampling budget
MethodClean ↑Random ↑Average ↑Δ
LingBot-VA(1/1)57.050.753.9—
Flash-WAM †(1/1)82.174.278.2—
Continued(1/1)82.376.179.2—
APM(1/1)84.377.580.9+1.7
LingBot-VA(2/2)88.386.087.2—
Flash-WAM(2/2)90.588.689.6—
Continued(2/2)89.687.788.6—
APM(2/2)90.487.188.8+0.1
LingBot-VA(4/4)90.490.090.2—
Flash-WAM(4/4)92.090.491.2—
Continued(4/4)90.589.189.8—
APM(4/4)92.590.391.4+1.6

Success rates (%). Local evaluations cover 50 tasks × 50 episodes per condition. Each student variant uses one checkpoint across all budgets. Δ is the gain over Continued, computed before rounding. † Flash-WAM’s one-step result is reported by DirectWAM; its two- and four-step results are evaluated locally. Full-budget LingBot-VA (25/50, reported): 92.9% Clean / 91.6% Random.

Better video predictions at one step.

Compared with continued Stage I distillation, APM increases single-step PSNR from 30.169 to 32.317 dB, alongside improvements in SSIM and LPIPS. The evaluation uses 500 shared contexts with ground-truth observation histories. Against Flash-WAM under each model’s deployment configuration, APM has higher PSNR; Flash-WAM has better SSIM and LPIPS.

Single-step gains in PSNR, SSIM, and LPIPS relative to continued distillation
Positive values indicate improvement over continued distillation. Video/action CFG scales are 5/1 for LingBot-VA and APM, and 1/1 for Flash-WAM.
Ground truth, LingBot-VA, Flash-WAM, and APM future-frame comparisons for Adjust Bottle and Pick Diverse Bottles
Single-step predictions for Adjust Bottle and Pick Diverse Bottles. Frames 1, 4, and 8 are shown with enlarged details from frame 4. These examples were selected post hoc from the 500 contexts based on APM’s improvements over Flash-WAM.

From predicted futures to physical actions.

Astribot S1 platform and physical execution stages for inserting a pen and folding a towel
Astribot S1 with dual 7-DoF arms. Each method is evaluated over 24 trials per task with matched initial states.

203.1 ms per control cycle

Without task-specific retraining, single-step APM improves success over the coarse-grid teacher on both tasks. Compared with the full eight-step teacher, it reduces latency from 4323.6 to 203.1 ms.

The task tradeoff remains visible: APM has lower success on PenHolder and higher success on FoldTowel than the full teacher. Inference runs on an NVIDIA A800 GPU with 80 GB of memory.

Real-robot task success and control-cycle latency
Method (video/action steps)PenHolder ↑FoldTowel ↑Latency ↓
Teacher (8/8)83.3%70.8%4323.6 ms
Coarse teacher (1/1)29.2%25.0%187.5 ms
APM (1/1)62.5%75.0%203.1 ms

Citation

@misc{liu2026anchoredpathmatching,
  title = {Distilling World Action Models for Efficient Bimanual Manipulation via Anchored Path Matching},
  author = {Jialong Liu and Junjie He and Yandu Chen and Bin Zhou and Heng Su and Peijun Tang and Jianan Wang and Changhao Chen},
  year = {2026},
  eprint = {2611.22592},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  doi = {10.48550/arXiv.2611.22592},
  url = {https://arxiv.org/abs/2611.22592}
}