TLDR: TrackVLA++ is a new Vision-Language-Action (VLA) model designed for embodied visual tracking, enabling robots to continuously follow moving targets. It addresses limitations of previous models by incorporating a spatial reasoning mechanism called Polar-CoT and a Target Identification Memory (TIM). Polar-CoT efficiently infers a target’s relative position, while TIM maintains a robust visual identity of the target over time, even during occlusions. This allows TrackVLA++ to achieve state-of-the-art performance in complex, crowded, and long-horizon tracking scenarios, demonstrating strong generalization in both simulated and real-world environments.
Embodied Visual Tracking (EVT) is a crucial capability for robots, enabling them to continuously follow moving targets in real-world applications like companion robots or service assistants. While recent advancements have allowed language-guided tracking in complex environments, existing methods often struggle with severe occlusions or when faced with similar-looking distractions. This is primarily due to a lack of explicit spatial reasoning and effective long-term memory.
A new research paper introduces TrackVLA++, a novel Vision–Language–Action (VLA) model designed to overcome these challenges. TrackVLA++ significantly enhances embodied visual tracking by integrating two key modules: a spatial reasoning mechanism and a Target Identification Memory (TIM).
Smarter Tracking with Polar-CoT Reasoning
At the heart of TrackVLA++’s reasoning capability is the Polar Chain-of-Thought (Polar-CoT) mechanism. Unlike traditional reasoning methods that might generate verbose textual plans or complex visual intermediates, Polar-CoT adopts a lightweight and efficient approach. It infers the target’s relative position—expressed as an angle and distance in the robot’s own coordinate system—and encodes this information into a compact ‘reasoning token’. This design is particularly efficient for dynamic scenarios like EVT and inherently supports multi-camera setups by avoiding the ambiguities often associated with bounding box predictions across different views. The model can even signal when a target is occluded or out of view with a special token, enhancing its robustness.
Remembering Targets with Target Identification Memory (TIM)
To ensure consistent target identification over long periods, especially during occlusions, TrackVLA++ introduces the Target Identification Memory (TIM). This module acts as a robust, persistent representation of the target’s visual identity. TIM employs a confidence-aware gating mechanism, meaning it only updates its memory state when the Polar-CoT mechanism confidently predicts the target’s presence. If the confidence is low, or if an ‘invalid’ token is generated (indicating occlusion or absence), the memory freezes, preserving the last reliable representation. This prevents memory corruption from distractors or drift during the target’s temporary disappearance, ensuring the robot can re-identify the target once it reappears.
Also Read:
- ContextVLA: Enhancing Robot Dexterity with Efficient Temporal Understanding
- MetaVLA: A Unified Training Approach for Generalist Robot Agents
State-of-the-Art Performance in Simulations and Real-World
Extensive experiments demonstrate that TrackVLA++ achieves state-of-the-art performance across various benchmarks. On the challenging EVT-Bench, TrackVLA++ significantly outperforms previous leading methods, particularly in ‘Distracted Tracking’ scenarios, showing improvements of 5.1% and 12% in success rate for egocentric and multi-camera settings, respectively. It also exhibits strong zero-shot generalization on the Gym-UnrealCV benchmark, successfully tracking targets for maximum episode durations even in unseen environments and distinguishing targets from identical distractors.
Beyond simulations, TrackVLA++ has been rigorously evaluated in real-world scenarios using a Unitree GO2 quadruped robot. In tasks involving obstacles, winding paths, and distractors, TrackVLA++ consistently outperformed its predecessor, TrackVLA, by substantial margins (14%, 7%, and 17% respectively in success rate). This highlights its remarkable robustness and practical applicability in dynamic and unpredictable real-world conditions.
An ablation study confirmed the individual contributions of both the Polar-CoT and TIM modules, with each contributing significantly to the overall performance gains. The research paper, available at https://arxiv.org/pdf/2510.07134, concludes that TrackVLA++ sets a new standard for embodied visual tracking by effectively integrating explicit spatial reasoning and long-horizon target memory, paving the way for more intelligent and reliable autonomous systems.
