qshou-coder/SpaceDreamer-Piper-Inference

SpaceDreamer Piper WAM

This repository contains the runtime artifacts for the Piper RGB-state WAM policy. Given one cam_high RGB frame, the current 14D dual-arm qpos, and one of four supported task IDs, the policy generates a 32-step action chunk.

Inference code and documentation: qshou-coder/SpaceDreamer, piper_inference branch

Artifacts

File	Purpose
policy.pt	Complete RGB DiT, ActionDiT, state encoder, and action normalizer
Wan2.2_VAE.pth	Wan2.2 VAE used to encode the live RGB frame
prompt_contexts.pt	Frozen UMT5 contexts for the four supported tasks
inference_config.yaml	Portable model and input/output contract
SHA256SUMS	Artifact integrity checksums

The policy checkpoint already contains the action normalization buffers. It does not require the Wan base DiT shards or the UMT5 encoder at runtime.

Input/output contract

• Image: uint8[H,W,3], from cam_high, explicitly marked RGB or BGR.
• State: raw, unnormalized qpos with shape [14].
• State/action order: left joints 1-6, left gripper, right joints 1-6, right gripper.
• Output: float32[32,14] absolute joint-position commands at 25 Hz.
• Sampling: joint RGB/action flow matching, 20 inference steps, shift 5.

Supported task IDs:

• assemble_battery_long
• battery_assemble
• pack_3_objects_plus
• stack_3_cups_gen

See the GitHub usage guide for the Python API and offline CLI.

Safety

The model only predicts actions. It does not enforce robot joint limits, velocity limits, observation freshness, collision avoidance, or emergency stop state. A robot-specific safety/control layer must validate every chunk before execution.

Upstream attribution

The included VAE is redistributed from Wan-AI/Wan2.2-TI2V-5B, which is licensed under Apache-2.0. The inference implementation also uses FastWAM components distributed with the accompanying GitHub code under MIT.