Embodied Video Foundation Model: Vidar & AnyPos

Vidar includes a video diffusion foundation model for video prediction and a masked inverse dynamics model (MIDM) for action regression. The video diffusion model is trained on 750K multi-view bimanual videos with test-time scaling (TTS) applied during testing, and it can adapt to new robot platforms with only 20 minutes of demonstrations with state-of-the-art performance and generalize to unseen tasks with strong semantic understanding.

Key Techniques

• Video Generation Model: Rectified flow models with internet-videos pre-training, emboided pre-training and fine-tuning on unified observation space.
• Masked Inverse Dynamics Model (MIDM): Inverse dynamics models often suffer from poor generalization due to the presence of background noise, texture biases, and visual distractions in high-dimensional observation, while MIDM can focus on task-relevent regions of the input frame via implicit mask prediction.
• Test-Time Scaling (TTS): We generate K candidate video trajectories using different random seeds. Then rank these trajectories using a pretrained evaluator (e.g., CLIP or a vision-language model) and select the one which scores the highest.

Vidar Predicted Video Samples

Evaluation of Vidar on the RoboTwin 2.0 benchmark

AnyPos is a robot-specific image-to-action model trained entirely on task-agnostic trajectories sampled by ATARA. It integrates two key techniques to enhance performance: Arm-Decoupled Estimation and Direction-Aware Decoder (DAD). Together, ATARA and AnyPos constitute a fully task-agnostic framework for training IDMs without goal supervision. By combining scalable unsupervised data collection with physically informed learning architectures, our approach demonstrates that task-agnostic action data can serve as a practical and powerful foundation for generalizable manipulation.

Our AnyPos method achieves 57.13% action prediction accuracy on the test set, which includes unseen skills and objects, surpassing previous approaches (naïve ResNet+MLP used in unipi, unisim, robodreamer, and susie) by 51%. In real-world robot replay tests, AnyPos-ATARA demonstrates 92.59% task success rate (as shown in the figure below), representing a 33% improvement over the human-collected dataset, and surpassing previous approaches by 44% (as shown in the Figure of "Overview of AnyPos").

ATARA: Automatically collect data of target robots! (4x speed)

Task-Agnostic Action

Standard VLA models learn temporally extended policies $p_{\theta}(\bm{a}_{T+1:T+t}|\bm{x}_{T-H+1:T},\mathbf{l})$ , where $\theta$ represents the parameters of the VLA policy, T is the current timestep and H denotes the history window size, mapping observation histories and language commands to action sequences. Given an expert dataset $D_{\text{expert}}$ , the training objective of VLAs is to maximize the likelihood:

Figure 3: The objective of VLAs.

Figure 4: Decompsing task-specific action.

Key Techniques

Figure: Methods of AnyPos. We obtain a task-agnostic training dataset covering the entire cubic workspace of dual robotic arms using ATARA.

• ATARA: As shown in the Figure of "Methods of AnyPos", naïve joint-space sampling often results in inefficient coverage of reachable states, redundant or degenerate motions (e.g., arms exiting the field of view), and frequent self-collisions. To address these limitations, we propose ATARA, a reinforcement learning framework that constructs a coverage-aware mapping from end-effector space to joint space. Therefore, it enables efficient task-agnostic data generation that preserves inherently encoded robot embodiment information and broad behavioral coverage, serving as a reusable prior for downstream policy learning.
• Arm-Decoupled Estimation: We observe that when estimating left arm joints, the model often attends to visual features of the right arm, and vice versa. To mitigate this, we isolate the input features per arm. We (1) split the left and right arms in the observation x, then (2) use two sub-networks estimate joint positions for each arm independently. Gripper poses are estimated by specialized networks. This decoupling reduces the visual hypothesis space, improves estimation accuracy, and enables specialization per arm.
• Direction-Aware Decoder (DAD): For action estimation networks, we choose DINOv2 with register (DINOv2-Reg) as the visual encoder, and uses three core componets to meet the high-precision requirement for action prediction: (1) Multi-Scale Dilated Convolutions (2) Deformable Convolutions (3) Angle-Sensitive Pooling