Enabling Robots to “Think While Acting”: LingBot-VA Paper Accepted at RSS 2026

Research Motivation

For robots, the difficulty lies not only in executing actions but also in understanding the resulting changes in the environment—for example, how the tabletop changes after picking up a cup or how object positions shift after pulling a drawer. The study introduces the ability for robots to first predict how the world will evolve and then decide on actions based on those predictions.

Technical Approach

We propose a causal world modeling framework for robot control and implement it as the world’s first open‑source autoregressive video‑action model, LingBot‑VA. The model continuously predicts environmental changes during task execution and generates the next action command, giving robots a human‑like “observe‑judge‑act” capability.

LingBot‑VA embeds causal relationships into its architecture: each prediction step relies only on prior observations and actions, respecting the forward‑in‑time nature of physical reality. The system uses a Mixture‑of‑Transformers (MoT) architecture that unifies video prediction and action generation within a single autoregressive diffusion framework, and it incorporates a closed‑loop inference mechanism that constantly ingests real‑world feedback to reduce error accumulation over long horizons.

Experimental Results

On the RoboTwin 2.0 benchmark covering 50 dual‑arm manipulation tasks, LingBot‑VA achieves average success rates of 92.0% in the Easy setting and 91.1% in the Hard setting. On the LIBERO benchmark it reaches 98.5%.

In real‑world evaluations involving six high‑difficulty challenges (long‑horizon, high‑precision, flexible and articulated object manipulation), LingBot‑VA adapts with only 50 real demonstration trajectories and improves overall success rate by more than 20 percentage points compared with the industry baseline π0.5, demonstrating strong data efficiency and generalization.

Open Source and Future Work

The paper’s acceptance at RSS 2026 marks international recognition of AntLingbo’s exploration of world‑model‑driven robot control. Model weights, training code, and inference scripts were released in January 2026 on the Magic‑Dock community, Hugging Face, and GitHub. The authors anticipate that this causal world‑modeling line will help robots move from command‑following toward deeper environment understanding, broader task generalization, and autonomous decision‑making.