Song

Model · Image to video

flux3-agibot-looped

Song

Looped Diffusion Transformer (Chng et al., 2026, arXiv:2609.40305) retrofit of Jinyan0924/flux3-agibot-videoft (itself a full fine-tune of black-forest-labs/flux-3-action-base): the 33 joint blocks are split 12 / 9 / 12, the middle 9 are run N = 4 times with head-wise gated attention, and every exit is trained with deep supervision. Task: k past head-camera frames + caption → 33 future frames at 15 fps, Val loss (EMA, N = 4, lower is better; k = number of context frames): r1 step 2500 k1 0.0787 / k9 0.0654; r2 step 4000 k1 0.0783 / k9 0.0651 (the step-4500 val pass was interrupted by the crash that ended the run, so the last val point is step 4000). Test-split PSNR / SSIM / LPIPS / FVD and…

Model · Robotics

qwen_robotics_nav_policy

Song

An image-in, path-out navigation policy for a slow walking robot: it looks at the last few front-camera images, reads a prompt that contains the goal ("Please walk towards the goal (4.0, -1.0)."), and returns the next 2 m of path as 8 waypoints. It is Qwen2.5-VL-3B with LoRA adapters and a small action head, trained by imitation on open navigation data. Code, data conversion and evaluation: Final-frame forecasting model (recommended): ff3lora4bdedup/ — Qwen/Qwen3.5-4B-Base + LoRA (rank 32) on the language model, action head MLP. Input: 5 past frames (1 Hz), the current frame and the frame 5 s ahead, plus one sentence saying who carries the camera (human / robot); output: 10 positions at 0.5…

Dataset · Robotics

habitat_hssd_pointgoal_nav_scenarios

Song

Image-in, trajectory-out navigation scenarios with 3D ground truth, generated in simulation with Habitat-Sim in the houses of HSSD (Habitat Synthetic Scenes Dataset, Khanna et al. 2023). Same schema and evaluator as the real-robot conversions JRDB); the CODa card documents every column, the coordinate frame, the point and map formats and the evaluation protocol. This set is the large training counterpart of those small real-world test sets. Code: - Input, the past. Rendered forward-camera images (640 x 480, 90° horizontal field of view, 1.2 m above the floor) of the 10 past steps and the current step, the agent's own past positions, and the - Ground truth, the future. The shortest navmesh…