TIC-VLA

University of California, Los Angeles (UCLA) · 2026-02

VLANavigationNot disclosedregression

A think-in-control VLA in which a fast action policy controls asynchronously, conditioned on the delayed state from slow VLM inference and explicit latency information (inference·elapsed time, ego-motion).

View architecture diagram in the gallery Paper ↗

Paper

TIC-VLA: A Think-in-Control Vision-Language-Action Model for Robot Navigation in Dynamic EnvironmentsarXiv 2602.02459

Overview

Category: Navigation
Type: VLA
Released: 2026-02
Organization: University of California, Los Angeles (UCLA)
Open weights: Yes
Parameters: Not disclosed (InternVL3-1B backbone, HF lists 0.9B; action expert size Not stated)

Architecture

Backbone LLM: Qwen2.5-0.5B · InternVL3-1B config
Vision encoder: InternViT-300M (built into InternVL3-1B, shared vision encoder)
Action head: Continuous 3-D action chunk MLP regression (integrated-trajectory smooth L1) (regression)
Chunk: 3-second prediction horizon (T=30 steps)
Control rate: 10 Hz (VLM 0.5 Hz)
Architecture (original summary): Shared vision encoder + InternVL3-1B VLM (slow reasoning 0.5 Hz, delayed last-layer KV cache projected by MLP) + Transformer action expert (learned action queries cross-attend to current visual tokens·delayed VLM cache·robot state·delay Δt·ego-motion Δp) → MLP outputs continuous 3-dim action chunks, executes the goal point of the integrated 3 s trajectory, 10 Hz

Training

Pretraining accelerators: 8 L40S (VLM SFT, DDP)
Pretraining duration: Not disclosed
Pretraining steps & batch: VLM SFT batch 2 per GPU (×8)·lr 2e-5, action expert batch 16 per GPU·lr 2e-4; 10 epochs·1,000 warmup steps (Table 8, no stage split); total steps not disclosed
GPU/TPU-hours: Not disclosed
Fine-tuning: PPO online RL: single L40S, 400 iterations (3 tasks×3 envs); time not disclosed
Data type: Real-robot teleoperation / simulation·synthetic
Data scale: SCAND 8.7 h + GND 11+ h + DynaNav 5.1 h (total 24.8+ h); CoT·instructions auto-annotated by GPT-5; sample·episode counts not disclosed
Main datasets: SCAND; GND; DynaNav (in-house Isaac Sim)
Tasks & evaluation: DynaNav (85 tests, warehouse · hospital · office · sidewalk) SR 55.29% · SPL 50.29 · collision rate 28.24% (SR 47.06% without RL; NaVILA 28.24%, Uni-NaVid 22.35%, point-goal NavDP 54.12%); real robot 4 tasks × 5 runs SR 0.85 (RTX 4060) · 0.75 (Orin NX) · 0.80 (A6000), DualVLN 0.50 · NaVILA 0.35
Robot: Real robot Unitree Go2 quadruped (front RGB); sim DynaNav Nova Carter wheeled·Boston Dynamics Spot quadruped

Inference

Inference device & speed: Action policy/VLM inference latency: RTX 4060 laptop (50W) 85.73/3430.73 ms, Jetson Orin NX (25W) 120.27/4831.73 ms, RTX A6000 32.70/1681.66 ms; sim (L40S) actions 10 Hz·VLM 0.5 Hz; memory not disclosed
Inference memory: Not disclosed
Weights bf16 (calc.): Not disclosed
Where it runs: Edge GPU (Jetson Orin NX·RTX 4060 laptop)

Reinforcement learning & base

Reinforcement learning: Used (VLA RL post-training) · After imitation learning, PPO online RL on the action policy only (Isaac Lab dynamic pedestrian environment, vision encoder·LM frozen)
RL evidence: https://arxiv.org/html/2602.02459v2
Base control: Wheeled·Quadruped (Waypoints → tracking controller) · Action policy outputs continuous action chunks → the robot tracks the goal point of the integrated short-horizon trajectory (in sim a controller generates smooth velocity commands (v, ω); real-robot low-level command format not stated); real robot Go2, sim Nova Carter·Spot

Notes

Key features: A think-in-control VLA in which a fast action policy controls asynchronously, conditioned on the delayed state from slow VLM inference and explicit latency information (inference·elapsed time, ego-motion). Injects 0–10 s inference delays during training and fine-tunes with PPO, maintaining 10 Hz control even with multi-second delays and reaching DynaNav SR 55.29%, ahead of prior VLAs (NaVILA 28.24%).
Notes: ICML 2026. Checkpoints·inference dataset·DynaNav·RL scripts released (HF handsomeYun/TIC-VLA dataset repo). Total parameters·training time not disclosed. Data hours are the sum of the three datasets' stated values (lower bound, since GND says 'over 11 hours'). SCAND is human teleoperation demonstrations (SCAND paper abstract); GND collection method unverified. Real-robot low-level command method not stated.
Verification: Corrections applied