MotionVLA

Injecting Geometric Motion into Vision-Language-Action Models

Should a VLA remember past frames—or the motion that connects them?

Shanglin Yuan^1,2 Weiheng Zhao^1,2 Xianda Guo^2,3 Wei Sui² Li Yu¹ Wenyu Liu¹ Xinggang Wang¹

¹ Huazhong University of Science and Technology ² D-Robotics ³ Wuhan University

Accepted to Conference on Robot Learning (CoRL), 2026

Paper Code

41 → 53 RoboTwin average SR

19 → 41 Touch SR

85.2 → 91.2 LIBERO-Long SR

Comparison of fragmented discrete 4D injection with MotionVLA's continuous trajectory-field interface, alongside RoboTwin results. — Discrete per-frame geometry can fragment motion; MotionVLA exposes a time-continuous, queryable history instead.

Motivation

Remember motion, not just more frames.

Long-horizon manipulation is temporally ambiguous: the same current image can require continuing, stopping, or advancing a subgoal. Adding history alone does not solve this when the injected evidence is geometrically inconsistent or temporally fragmented.

MotionVLA converts a short, strictly past-only head-camera window into compact trajectory-field tokens. The policy retrieves only the motion evidence relevant to its current observation, preserving recent physical progress without flooding the VLA context with raw video.

A useful memory should tell the policy how it arrived here, not merely show it more images.

Method

A queryable motion-history interface.

The current frame stays in the VLA perception stream. Only earlier frames enter the frozen trajectory extractor.

MotionVLA architecture: a frozen trajectory extractor encodes history, current visual tokens query the resulting motion history, and trajectory grounding supervises fusion into the VLA. — MotionVLA architecture from the paper.

Past-only history

At time t, the current frame is excluded. The default 10 FPS history selects five frames with stride two and pads only from earlier observations.

Compact 4D tokens

A frozen TraceAnything encoder runs once on the sampled window. MotionVLA keeps final decoder tokens rather than expanding the dense trajectory field inside the policy.

Decouple

Current head-camera tokens act as queries; normalized historical 4D tokens are keys and values. Cross-attention retrieves motion evidence aligned to what the policy sees now.

Ground and recouple

Historical EEF reconstruction keeps retrieved tokens dynamic. Recouple optionally mixes them back with current multi-view tokens before action generation.

Paper-reported results

Results on RoboTwin and LIBERO.

Parenthetical deltas compare MotionVLA with π₀.

RoboTwin2.0

Success rate (%)

LIBERO

Success rate (%)

MotionVLA Successful simulation rollout · 2× speed

π₀ Failed simulation rollout · 2× speed

Motion consistency

Path efficiency.

Path efficiency (PE) divides policy EEF travel by expert travel. Lower is better; 1.0 matches expert path length.

LIBERO

Success and path efficiency

LIBERO path efficiency on Goal and Long
Method	Goal		Long
Method	SR ↑	PE ↓	SR ↑	PE ↓
π₀	95.8	1.23	85.2	1.19
4D-VLA*	88.8	1.88	72.6	2.06
4D-VLA	89.8	1.36	78.6	1.32
MotionVLA	96.2	1.05	91.2	1.10

PE is averaged only over successful episodes and is shown with SR. 4D-VLA* uses pseudo RGB-D predicted by Depth Anything 3.

Real Robot Experiments

Real-robot results.

Three temporally demanding tasks, with more than 100 demonstrations per task.

Agilex Piper

Success rate and completion steps

Real-world Agilex Piper results with tasks as columns and methods as rows
Method	Pick & Place		Ranking		Touching		Average
Method	SR	Steps	SR	Steps	SR	Steps	SR	Steps
Demonstration	—	119	—	351	—	261	—	244
π₀	80.0	118	6.1	558	0.0	—	28.7	338
MotionVLA	92.5	119	18.2	443	15.0	285	41.9	282

Steps equal completion time × 10, rounded to the nearest integer. Unavailable task-level steps are excluded from each method's average; π₀ has no Touching completion-step value.

Citation

@inproceedings{yuan2026motionvla,
  title     = {MotionVLA: Injecting Geometric Motion into Vision-Language-Action Model},
  author    = {Yuan, Shanglin and Zhao, Weiheng and Guo, Xianda and Sui, Wei and Yu, Li and Liu, Wenyu and Wang, Xinggang},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026}
}