UFACTORY-Developer/ufactory_genesis_rl

UFACTORY Genesis Reinforcement Learning Checkpoints

Pretrained reinforcement learning policy checkpoints for ufactory_genesis, built on Genesis and rsl-rl PPO.

---

1. Reach (reach/)

Bare-arm target reaching for 5 UFACTORY manipulators (xarm5, xarm6, xarm7, uf850, lite6):

• Initial flange TCP pose: (300, 0, 300) mm, RPY = (180°, 0°, 0°)
• Success criterion: strictly < 5 mm terminal distance (success_distance_m = 0.005)
• Control rate: 50 Hz (20 ms step, 5 s episode horizon), normalized joint-position increments ([-1, 1] -> max 0.01 rad/step)
• Training: rsl-rl PPO ([256, 256, 128] ELU MLP with Beta distribution), 4096 parallel environments, 300 iterations, seed 7

Benchmark & Holdout Validation

Evaluated on two independent holdout seeds (17000 and 17001, 512 random targets each):

Robot (--robot)	DOF	Obs / Act Dim	Target Workspace (m)	Checkpoint	Holdout Success Rate (17000 / 17001)	Mean Terminal Distance (17000 / 17001)
xarm5	5	21 / 5	x 0.15–0.55, y -0.30–0.30, z 0.05–0.50	model_299.pt	99.6% / 99.8%	4.48 / 4.50 mm
xarm6	6	24 / 6	x 0.15–0.55, y -0.30–0.30, z 0.05–0.50	model_299.pt	100% / 100%	4.23 / 4.24 mm
xarm7	7	27 / 7	x 0.15–0.55, y -0.30–0.30, z 0.05–0.50	model_299.pt	99.8% / 99.2%	4.39 / 4.52 mm
uf850	6	24 / 6	x 0.15–0.55, y -0.30–0.30, z 0.05–0.50	model_299.pt	99.6% / 99.6%	4.26 / 4.34 mm
lite6	6	24 / 6	x 0.15–0.35, y -0.20–0.20, z 0.05–0.35	model_299.pt	99.8% / 99.6%	4.66 / 4.71 mm

---

Download & Usage

Run from the root of the ufactory_genesis repository:

# Download all 5 Reach checkpoints into outputs/rl/reach/
hf download UFACTORY-Developer/ufactory_genesis_rl --include "reach/**" --local-dir outputs/rl

# Or download a single robot (e.g. xarm6)
hf download UFACTORY-Developer/ufactory_genesis_rl --include "reach/xarm6-reach/**" --local-dir outputs/rl

Real-time GLB Inference

# Single arm
python -m examples.rl.reach.visualize \
  --checkpoint outputs/rl/reach/xarm6-reach/seed-7/model_299.pt

# Four arms
python -m examples.rl.reach.visualize \
  --checkpoint outputs/rl/reach/xarm6-reach/seed-7/model_299.pt -b 4

Quantitative Evaluation

python -m examples.rl.reach.evaluate \
  --checkpoint outputs/rl/reach/xarm6-reach/seed-7/model_299.pt \
  -B 128 --episodes 512 --seed 17000 \
  --min-success-rate 0.95 --max-mean-distance-m 0.010 \
  --quiet-episodes --summary-json outputs/rl/reach/xarm6-reach/seed-7/evaluation-17000.json

See the full guides in the code repository:

• English Guide (examples/rl/reach/README.md)
• Chinese Guide (examples/rl/reach/README_zh.md)