← Paper Digest

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

Jialong Liu, Junjie He, Yandu Chen, Bin Zhou, Heng Su, Peijun Tang, Jianan Wang, Changhao Chen

November 2026

RoboTwin 2.0 simulation: a gripper places a can in the basket.

TL;DR Distills the LingBot-VA world action model into one student that generates future video and bimanual actions in 1, 2 or 4 denoising steps. One-step APM reaches 80.9% average success on RoboTwin 2.0 and, on a real Astribot S1, cuts control-cycle latency 21.3x versus the eight-step teacher.

What they do

World action models (WAMs) generate future video and robot actions together by iterative denoising, which is slow: the eight-step LingBot-VA teacher needs 4.3 s per control cycle on the real robot. Anchored Path Matching (APM) distills it into a single student checkpoint that runs at 1, 2 or 4 steps, with the step count chosen at deployment and no teacher needed at inference.

How it works

  • Stage I. LingBot-VA is distilled with AnyFlow into a finite-interval transition map that jumps directly from one noise level to a target level. Teacher targets come from eight integration steps.
  • Student-reached states. Coarse steps drift off the teacher’s path, so Stage II rolls out the student at 1, 2 or 4 steps and supervises a state it actually reaches. The teacher continues from that state to zero noise; its endpoint is the anchor target and its first intermediate state the local-transition target.
  • Anchor-priority correction. The backbone is frozen and separate LoRA adapters for video and action are trained, keeping the base distillation loss. Where the local-transition gradient opposes the anchor gradient, that component is projected out within each modality.

Results

  • RoboTwin 2.0 (50 tasks, 2,500 episodes per condition, trained on Clean data only): APM averages 80.9%, 88.8% and 91.4% success at 1, 2 and 4 steps, against 79.2%, 88.6% and 89.8% for continued base distillation with matched parameters and training steps.
  • Single-step video PSNR rises from 30.169 to 32.317 dB over continued distillation; Flash-WAM keeps better SSIM and LPIPS.
  • Real robot (24 trials per task): one-step APM succeeds 62.5% on PenHolder and 75.0% on FoldTowel, against 29.2% and 25.0% for the one-step teacher and 83.3% and 70.8% for the eight-step teacher, at 203.1 vs 4323.6 ms per cycle.

Limitations

Real-robot tests cover two tasks on one platform, with inference on an A800 GPU. At two steps the gain is 0.1 points, with Random success 0.6 points lower. Local-transition supervision without projection does not improve closed-loop success, and projection raises local-transition error.

Videos

RoboTwin 2.0 simulation: a gripper picks up a phone and brings it to the phone stand.
RoboTwin 2.0 simulation: a gripper places a mouse on a yellow mouse pad.

Figures

From Figure 1: Coarse few-step student transitions can drift from the dense eight-step teacher trajectory and change ...
From Figure 1: Coarse few-step student transitions can drift from the dense eight-step teacher trajectory and change the remaining prediction. With the same initial condition and four steps, the naive student fails and APM succeeds; squares are video states and circles action states.
From Figure 2: (a) One teacher continuation from the student-reached state supplies both the restart endpoint and the...
From Figure 2: (a) One teacher continuation from the student-reached state supplies both the restart endpoint and the next local target; the gray path is the Original anchor used in the ablation. (b) Endpoint and short-transition queries share that state, and anchor-priority correction combines their gradients with the base objective while the backbone stays frozen. (c) At deployment the student generates video and actions in K steps without the teacher.
From Figure 4: Single-step video quality gains over continued base distillation in PSNR, SSIM and LPIPS on 500 shared...
From Figure 4: Single-step video quality gains over continued base distillation in PSNR, SSIM and LPIPS on 500 shared contexts; positive values are improvements. APM is the Restart anchor plus field supervision.
From Figure 6: The Astribot S1 with an onboard camera, two 7-DoF arms and grippers, and the two real tasks: PenHolder...
From Figure 6: The Astribot S1 with an onboard camera, two 7-DoF arms and grippers, and the two real tasks: PenHolder (reach, grasp, insert) and FoldTowel (initial state, lift, fold).
Abstract

World Action Models (WAMs) 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 and short-transition predictions from states reached by the student. From each state, we continue the teacher to zero noise and use its endpoint and first intermediate state as targets. Modality-routed adapters learn to match both targets from the same starting state. When the two gradients conflict, anchor-priority updates preserve the endpoint gradient and remove the opposing component of the local gradient within each modality. One checkpoint supports 1-, 2-, and 4-step joint generation. The sampling budget can be chosen at deployment without retraining or teacher queries. On RoboTwin 2.0, we train only on Clean demonstrations and evaluate directly on the domain-randomized Random setting. At one step, APM improves closed-loop success over continued base distillation at matched inference budgets by 2.0 percentage points on Clean and 1.4 on Random. Video peak signal-to-noise ratio (PSNR) rises from 30.169 to 32.317 dB, with gains in structural similarity (SSIM) and learned perceptual image patch similarity (LPIPS). On two real-robot tasks, APM improves success over the one-step teacher and reduces control-cycle latency by $21.3\times$ relative to the eight-step teacher.Project page: https://jialonglong.github.io/Anchored-Path-Matching/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.