WorldFoundry
English中文

AHA-WAM

Method overview: AHA-WAM architecture and attention mask. AHA-WAM decouples world planning and action execution into a slow video-DiT planner and a fast action-DiT executor. The video branch is trained with a fully causal mask to.

World-action model for RoboTwin 2.0: video + action experts, six-frame history, 16×14 chunks. MIT license. A100 Workspace validated; not a closed-loop score.

Edit this pageGitHubMarkdown

AHA-WAM is a world-action model for manipulation: a video expert and an action expert share a mixture cache so the policy can use predictive world modeling at infer time. WorldFoundry runs the exact video/action experts, mixture-cache editor, six-frame history, two-phase chunk inference, Wan VAE, and UMT5 encoder in process. Trainers, simulators, and deployment servers are excluded. Pipeline: AHAWAMPipeline. Aliases aha-wam and aha-wam-robotwin resolve here.

Official paper teaser: Overview of AHA-WAM. AHA-WAM connects past observations, future-oriented world planning, and fast closed-loop action execution: a slow world planner maintains reusable memory and planning context, while a fast action.
Official paper teaser: Overview of AHA-WAM. AHA-WAM connects past observations, future-oriented world planning, and fast closed-loop action execution: a slow world planner maintains reusable memory and planning context, while a fast action.
Method overview: AHA-WAM architecture and attention mask. AHA-WAM decouples world planning and action execution into a slow video-DiT planner and a fast action-DiT executor. The video branch is trained with a fully causal mask to.
Method overview: AHA-WAM architecture and attention mask. AHA-WAM decouples world planning and action execution into a slow video-DiT planner and a fast action-DiT executor. The video branch is trained with a fully causal mask to.

The official paper is on arXiv 2606.09811. Project: serene-sivy.github.io/aha-wam. Code: Serene-SIVY/AHA-WAM. License: MIT. Recorded policy: SereneC/AHA-WAM-RoboTwin2.0 (robotwin_ahawam.pt). VAE and UMT5 come from DiffSynth-Studio/Wan-Series-Converted-Safetensors.

The released RoboTwin 2.0 checkpoint completed real A100 Workspace inference from three ALOHA views plus a 14-D state and returned a finite 16×14 open-loop chunk. Catalog runner is still unset — that is numerical checkpoint validation, not a closed-loop simulator score.

Build your run

Pipeline
AHAWAMPipelineworldfoundry.synthesis.action_generation.ahawam.pipeline
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  ahawam \
  --pipeline.task-profile default \
  --json

What it is for

AHA-WAM RoboTwin chunk from three ALOHA views plus state.

Launch worldfoundry-eval run ahawam.

Score on RoboTwin 2.0 — the released checkpoint is trained for RoboTwin 2.0.

Runtime contract

Required: image and prompt; optional: model_id, state, preferred_cameras, image_size, configured_action_horizon, action_chunk_size, and default_chunks_per_video_prefill; artifacts: action_trace and ahawam_action_trace.json.

KnobRecorded value
Image size384×320
Action horizon64
Chunk size16
Video prefill chunks2 (then KV refresh; up to six earlier prefill entries kept)

Run in WorldFoundry

Status: Verified. Unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8). Stage SereneC/AHA-WAM-RoboTwin2.0, DiffSynth-Studio/Wan-Series-Converted-Safetensors, and Wan-AI/Wan2.1-T2V-1.3B locally — missing files fail at load, nothing is downloaded. Run worldfoundry-eval run ahawam. Catalog VRAM is unset.