ManiFlow Real-Robot Baseline (plastic / pill)

ManiFlow DiTX image policy on Inspire-G1 real-robot Stage-2 data.

Org: Humantwin
Repo: Humantwin/maniflow

Contract

• Action chunk [30, 41] = Body29 (GMT) + Hand12 (Inspire) @ 30 Hz
• Obs: RGB [B,1,3,H,W], state [B,1,41]
• Vision: R3M + DiTX (n_layer=12, n_emb=768, visual_cond_len=1024)
• Checkpoint: 15k / 30k

Layout

plastic/
  step15000.ckpt
  latest.ckpt          # symlink → step15000
  norm_stats.json
pill/
  step15000.ckpt
  latest.ckpt
  norm_stats.json

Train loss @ 15k (approx)

Task	train loss	val_loss@15k
plastic	~0.05	(see local logs; val higher)
pill	~0.08	lower than plastic

Load

import torch
from maniflow.deploy.maniflow_robot_worker import ManiFlowRobotWorker

w = ManiFlowRobotWorker("plastic/step15000.ckpt", device="cuda", use_ema=True)
w.load()
chunk, ms = w.infer({"rgb": rgb, "state": state})  # chunk: [30,41] raw