Unifying Velocity and Endpoint Predictionfor Robot Manipulation

Zhengtu Liang1,*Qiwei Liang1,*,§Tao Lin2Shihang Weng3Boyang Cai1Xiangyu Chen1
Yikeng Chen1Jiaming Liang1Bohan Tan1Yu Yan1Renjing Xu1,‡

* Equal contribution§ Project leader‡ Corresponding author

1 The Hong Kong University of Science and Technology (Guangzhou)

2 Beijing Jiaotong University 3 Shenzhen University

the UniVEP teaser
A study of action prediction. From prediction targets and their noise-dependent behavior to efficient robot control. Original PDF

Prediction targets

Two predictions, one path.

A generative policy turns Gaussian noise into an action chunk, conditioned on observations, language, and robot state. What should its action network predict along that path?

Predict the endpoint

X-prediction estimates the complete clean action chunk. The sampler uses this endpoint estimate to move the current noisy action toward it.

Predict the velocity

V-prediction estimates the velocity along the noise-to-action path. The sampler integrates these local directions over successive evaluations.

zt = (1 − t)ε + tx

t runs from noise to clean action; σ = 1 − t is the remaining noise. Here velocity describes generation, rather than the robot’s physical joint speed.

endpoint and velocity prediction on the same path
Endpoint and velocity prediction on the same path. X estimates the clean endpoint, while V estimates the sample’s local direction. Gray segments mark each predictor’s endpoint error. Original PDF

One more distinction: prediction versus supervision

The study compares V, X, and X–V. X–V predicts the same endpoint as X and uses the same inference update; its training loss emphasizes low-noise endpoint errors with the weight 1/max(σ, 0.05)². This lets us separate the network’s output from the space in which it is supervised.

Action geometry: what structure does the endpoint offer?

Across 50 RoboTwin tasks, clean action chunks concentrate their variation in fewer directions than velocity targets or Gaussian noise. Their mean participation ratio is 2.47, compared with 10.84 for velocity and 35.81 for noise.

the structure of action, velocity, and noise targets
Participation ratio, 99% energy rank, and TwoNN, averaged equally across tasks. Original PDF
Further geometry checks
the supplementary action-geometry diagnostics
Additional energy thresholds, local tangent variation, and residual measurements from the paper. Original PDF

Performance across inference budgets

Why prediction targets matter

Algebraically related predictions need not learn equally well. We first compare V, X, and X–V under matched training within each policy, then vary the number of network evaluations.

Single-target policies on RoboTwin Clean

Mean success across 50 tasks · the same trained weights at every budget

Hover or tap a point for exact values. Select a legend item to isolate a target. Focus a chart and use the left and right arrow keys to inspect budgets.

Mean ± SD across three training seeds. Each panel uses its own success-rate scale.

1–2 evaluations

X–V leads at low budgets

StarVLA
55.56%
LiLa-WAM Full
80.12%

With one evaluation, X–V reaches these success rates on RoboTwin Clean. It also leads V and X at two evaluations in both policies.

1 to 16 evaluations

V gains as the budget grows

35.92% → 53.84%

StarVLA’s success with V rises by 17.92 percentage points. Direct X moves from 40.52% to 42.52%. At eight evaluations, LiLa-WAM V and X–V reach 82.84% and 82.56%.

One-evaluation results for all three targets
RoboTwin Clean success at one evaluation (%)
PolicyVXX–V
StarVLA35.9240.5255.56
LiLa-WAM Full67.4073.3680.12
Architecture and pretraining change this ranking

FastWAM’s velocity-pretrained configuration favors V at its native ten-call budget: 73.24% success, compared with 65.88% for X and 34.20% for X–V. Resetting only the action expert raises X–V to 64.48%, while V remains strongest at 71.56%. The pretrained video pathway is retained.

These native-budget runs use separate rollouts from the inference-budget sweep. Within each policy, architecture, data, and update budgets are matched. On RoboTwin, StarVLA and FastWAM use 50 Clean demonstrations per task; LiLa-WAM Full uses 50 Clean + 500 Randomized demonstrations.

Could the gaps come from what each target predicts well at different levels of noise?

Prediction behavior across noise stages

What changes as the noise fades?

The task-success curves suggest different strengths. We probe predictions at matched observations and noise levels to distinguish recovering an action from noise from refining one that is already nearly correct.

prediction behavior across noise stages
Prediction behavior across noise stages. (a) V/X output-variance ratios from pure noise. (b) Endpoint MSE relative to the corrupted input at σ = 0.01. (c) Recovery across noise levels on StarVLA Randomized. Intervals and bands are the original 95% bootstrap intervals. Original PDF

High noise

X recovers more stable endpoints

120.74×

On StarVLA Clean, endpoints reconstructed from V have 120.74 times the variance of X-predicted endpoints. X is more stable across all six policy and split settings.

Near-clean actions

V makes a useful correction

0.334×

On StarVLA Randomized at σ = 0.01, V leaves 0.334 times the input mean squared error. Pure X leaves 52.1 times the input error.

Why V helps near the endpoint

x̂V − x = σ(v̂ − v)

The remaining noise σ scales the velocity prediction error. As σ approaches zero, that error has a smaller effect on the reconstructed endpoint. The low-noise comparison above uses Pure X, with direct endpoint supervision.

How the noise probes are measured

Each probe uses one network evaluation, with conditions and checkpoints fixed. The study uses 2,004 matched observation anchors across 12 tasks per split and eight matched noise draws per anchor, on Clean and Randomized scenes for StarVLA, FastWAM, and LiLa-WAM Full. It measures endpoint variance and reconstruction error; the preceding budget sweep measures closed-loop task success.

This suggests using X to recover the action from heavy noise, then V to refine it near the endpoint.

From the observations to a policy

UniVEP: learning both predictions.

UniVEP brings endpoint and velocity prediction into a shared action network. Training supervises both modes; each inference step evaluates the mode selected for the current noise level.

the UniVEP framework: joint training and noise-dependent inference
Joint training and noise-dependent inference. The two action streams share the DiT and VLM conditioning. Their mask separates the streams; consistency aligns the clean actions implied by X and V. Original PDF

Joint supervision and consistency

A mode embedding selects X or V in the shared network. Both see the same noisy action, noise level, and policy conditions. Velocity and endpoint regression train each output; consistency aligns the endpoint predicted by X with the endpoint reconstructed from V.

UniVEP uses a bounded velocity-weighted endpoint loss. UniVEP–X uses a flat endpoint loss; both retain velocity supervision and the same consistency objective. The weighting interacts with FastWAM’s velocity-pretrained components.

A noise-aware switch

σ* = 0.5
Clean action · σ = 0High noise · σ = 1
X
Endpoint recovery

σ ≥ σ*: select the endpoint mode.

On the native grids, both one- and two-call inference use X only. Their gains come from joint training. At larger budgets, later evaluations can use V, depending on the policy’s time grid.

Evaluating the joint policy

Success in simulation.

We now evaluate the jointly trained policy on the same manipulation benchmarks and inference budgets.

61.36%

StarVLA with one call

Above sixteen-call X–V at 59.96% on RoboTwin Clean.

85.96%

LiLa-WAM with two calls

Above sixteen-call X–V at 82.60% on RoboTwin Clean.

Across manipulation benchmarks

Each policy at its native budget · success (%)

Success across inference budgets

Original paper curves
success across inference budgets in the original paper
The original curves include the three single-target recipes, UniVEP, and UniVEP–X where reported. One- and two-call gains come from the jointly trained endpoint predictor; later evaluations can use velocity refinement, depending on the policy’s time grid. Original PDF
Numerical results and evaluation settings

Within-policy controls: architecture, training data, and update budget are fixed. StarVLA and FastWAM use 50 Clean demonstrations per task; LiLa-WAM Full uses 50 Clean + 500 Randomized demonstrations. Mean ± SD is shown where the main table reports three training seeds. The native-budget table and the NFE sweep report comparisons within each policy.

Comparisons are controlled within each policy. LiLa-WAM Full uses Randomized training demonstrations; StarVLA evaluates Randomized scenes using a Clean-trained policy. Standard LIBERO is near saturation, and FastWAM’s velocity-pretrained configuration remains an important exception.

On the real robot

Dual-arm manipulation, in practice.

A multitask StarVLA policy controls a 14-DoF ARX A7 with an egocentric RealSense D405 camera. All recipes use the same four-call controller and 50 evaluation trials per task.

Drawer Cube Place

Drawer Cube Place · ARX A7 · open the drawer and place the cubeDownload video
the real-robot setup and three manipulation tasks
Real-world setup and three dual-arm tasks. (a) ARX A7 and RealSense D405. (b) Drawer Cube Place. (c) Pick Two Cube. (d) Grab Stick. Original PDF
58.67%

Mean real-robot success
+9.3 pp over the strongest single-target baseline

Gains on all three tasks.

The largest gain is on Drawer Cube Place: 38% to 52%. Grab Stick improves from 76% to 84%, and Pick Two Cube from 34% to 40%.

50 demonstrations per task · 20,000 joint updates · four Euler steps / four network calls · same action interface and controller.

Exact success rates
StarVLA real-robot success rates in percent
RecipeDrawer
Cube Place
Pick
Two Cube
Grab
Stick
Mean
V24323831.33
X34344638.00
X–V38347649.33
UniVEP52408458.67

Success on the real robot

StarVLA · four calls per action chunk · 50 trials per task

All panels share a 0–100% scale and the same four-call inference budget. Hover, tap, or focus a bar for exact success rates. The expandable table retains every reported value.

Inference efficiency

Fewer calls, lower latency.

At matched policy and NFE, each step evaluates one mode and retains the counted inference FLOPs of a single-target policy. The larger wall-clock savings come from generating effective actions with fewer calls.

LiLa-WAM Full · RoboTwin Clean

Two evaluations beat sixteen.

Higher success and 77.1% lower latency than the strongest sixteen-step single-target baseline.

Success rate

+3.36pp
X–V at 16 evaluations: 82.60 percent. UniVEP at 2 evaluations: 85.96 percent. The vertical axis starts at zero.050100Success (%)82.60%85.96%X–V16 evaluationsUniVEP2 evaluations

Inference latency

−77.1%
X–V at 16 evaluations: 100.98 milliseconds. UniVEP at 2 evaluations: 23.16 milliseconds. The vertical axis starts at zero.04080120Latency (ms)100.98 ms23.16 msX–V16 evaluationsUniVEP2 evaluations

Batch 1 · single RTX 4090 · latency through delivery of a complete action chunk to the CPU. At matched policy and NFE, counted inference FLOPs are unchanged. Joint training has separate training cost.

The same comparison holds for StarVLA: one-call UniVEP reaches 61.36% success at 61.59 ms, versus 59.96% at 191.32 ms for sixteen-call X–V, a 67.8% latency reduction.

Timing protocol and training cost

Latency is measured through delivery of a complete action chunk to the CPU, at batch size one on an RTX 4090, after 20 warm-up calls and over 300 synchronized timings. Counted FLOPs cover dominant matrix multiplications, convolutions, and attention operations.

The inference statement is separate from training. In the measured StarVLA setup, joint training adds 6.93% counted training FLOPs and 15.57% per-update time.

Resources

Paper and citation.

Find complete benchmark results, ablations, noise-probe measurements, training details, and compute accounting in the paper.

Read the full paper
BibTeX
Download .bib
@misc{liang2026univep,
  title  = {UniVEP: Unifying Velocity and Endpoint Prediction
            for Robot Manipulation},
  author = {Liang, Zhengtu and Liang, Qiwei and Lin, Tao and
            Weng, Shihang and Cai, Boyang and Chen, Xiangyu and
            Chen, Yikeng and Liang, Jiaming and Tan, Bohan and
            Yan, Yu and Xu, Renjing},
  year   = {2026},
  eprint = {2611.26103},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2611.26103}
}

Paper figure

Open original PDF

Pinch to zoom on touch devices · Esc to close