RoboHP/metaworld_mt50_act

🤖 ACT Model Trained on Meta-World MT50

This repository provides an ACT model trained using the RoboHP/metaworld_mt50_v3_50d_512r dataset for 200K steps.

🧠 Training Details

• Training script: Provided by LeRobot

• Command:

  python train.py --dataset.repo_id=RoboHP/metaworld_mt50_v3_50d_512r \
      --policy.type=act \
      --batch_size 64 --num_workers 16 \
      --steps 200000
  

🧪 Evaluation

• Environment: AoqunJin/Metaworld - MT50-V3

• Camera view: corner4

  ⚠️ Note: You must flip the image and corresponding actions to match the original dataset format. See evaluation code (WIP) for details.

• Trials per task: 25

📊 Results

• For task difficulty classification, see: paper

Category	Task	Success Rate
Easy	Avg.	0.846
	Button Press	1.00
	Button Press Topdown	1.00
	Button Press Topdown Wall	1.00
	Button Press Wall	1.00
	Coffee Button	0.96
	Dial Turn	0.92
	Door Close	1.00
	Door Lock	0.24
	Door Open	0.92
	Door Unlock	0.96
	Drawer Close	1.00
	Drawer Open	1.00
	Faucet Close	1.00
	Faucet Open	1.00
	Handle Press	1.00
	Handle Press Side	0.96
	Handle Pull	0.20
	Handle Pull Side	0.00
	Lever Pull	0.48
	Plate Slide	1.00
	Plate Slide Back	1.00
	Plate Slide Back Side	1.00
	Plate Slide Side	1.00
	Reach	0.64
	Reach Wall	0.64
	Window Close	0.96
	Window Open	1.00
	Peg Unplug Side	0.80
Medium	Avg.	0.665
	Basketball	0.36
	Bin Picking	0.96
	Box Close	0.52
	Coffee Pull	0.76
	Coffee Push	0.96
	Hammer	0.36
	Peg Insert Side	0.48
	Push Wall	0.80
	Soccer	0.32
	Sweep	1.00
	Sweep Into	0.80
Hard	Avg.	0.687
	Assembly	0.92
	Hand Insert	0.64
	Pick Out of Hole	0.40
	Pick Place	0.72
	Push	0.80
	Push Back	0.64
Very Hard	Avg.	0.728
	Shelf Place	0.56
	Disassemble	0.76
	Stick Pull	0.60
	Stick Push	1.00
	Pick Place Wall	0.72

Overall Average Success Rate: 0.731