Skip to main content
Back to timeline
arXivSource publication:

UniVEP splits robot action prediction by noise level: one network pass reaches 61.36% on RoboTwin, beating a single-target baseline's sixteen passes at 59.96%

Synopsis

Flow-matching robot policies usually predict one target throughout action generation, yet recovering an action from noise and polishing a nearly clean one are different problems. UniVEP trains endpoint prediction and velocity prediction inside one action network, aligns them with a consistency constraint, and samples with endpoints at high noise before switching to velocity for low-noise refinement, still running the network once per step. On RoboTwin, StarVLA with UniVEP reaches 61.36% success in one evaluation, above the strongest single-target baseline's 59.96% at sixteen; LiLa-WAM with UniVEP reaches 85.96% in two evaluations, above the single-target sixteen-step 82.60%, cutting action-generation latency by 77.1%.

Source-provided article image: UniVEP: Unifying Velocity and Endpoint Prediction for Robot Manipulation
arXiv

Interpretation

A robot flow-matching policy's action network need not learn a single prediction target across the whole generation process; it can divide the work by noise level, predicting the clean action directly at high noise and the velocity for local correction at low noise. Flow-matching policies have typically predicted the velocity field that transports noise toward data throughout, or predicted the clean action throughout; the two targets are analytically interchangeable but pose different learning problems. On StarVLA and LiLa-WAM, velocity-weighted endpoint prediction (X–V) reaches 55.56% and 80.12% with one evaluation, above direct endpoint prediction at 40.52% and 73.36% and velocity prediction at 35.92% and 67.40%; velocity prediction gains more from added evaluations, rising from 35.92% to 53.84% at sixteen on StarVLA.

UniVEP has one action transformer emit both predictions through a mode embedding, supervises velocity, endpoint and their consistency during training, and calls only one mode per step at inference according to noise level. Speeding up generation has usually relied on consistency distillation or a separately trained student model, whereas this keeps the existing policy and changes only the prediction target and supervision. The consistency term aligns the directly predicted endpoint with the endpoint reconstructed from velocity; at native RoboTwin budgets it raises success by 4.04 percentage points on StarVLA, 3.36 on LiLa-WAM and 7.08 on FastWAM.

The gains at one or two evaluations come from the jointly trained endpoint predictor itself, not from switching modes during inference. On these sampling grids the one- and two-step samplers use endpoint prediction throughout, so the improvement cannot be attributed to an endpoint-to-velocity switch. One-step StarVLA UniVEP reaches 61.36%, above one-step X–V at 55.56% and sixteen-step X–V at 59.96%; two-step LiLa-WAM UniVEP reaches 85.96%, above the best single-target results at eight, ten and sixteen steps, 82.84%, 83.48% and 82.60%.

Perspective

The demonstrated scope is three simulated benchmarks, LIBERO, LIBERO-Plus and RoboTwin 2.0, plus three tasks on a 14-DoF ARX A7 dual-arm platform with 50 demonstrations per task and 50 trials per task; on the real tasks UniVEP averages 58.67% success, 9.3 percentage points above the strongest single-target baseline, X–V. A researcher or engineer working on flow-matching manipulation policies can use this to judge how prediction target pairs with evaluation budget; the switching threshold is currently selected by a StarVLA success sweep at sixteen evaluations.

FastWAM is an explicit counterexample: its video backbone and action expert are both pretrained for velocity prediction, and at ten evaluations velocity prediction reaches 73.24% versus 57.00% for UniVEP; training only the action expert from scratch raises X–V from 34.20% to 64.48%. Joint training also carries a training-time cost, adding 6.93% counted FLOPs and 15.57% per-update time over velocity prediction for StarVLA on one A800; the switching threshold is currently chosen by a task-success sweep, and learning the switching rule from policy-specific evidence is listed as future work.