wpy123co/so101-wall-oss-fullft

WALL-OSS-0.5 Full Fine-tune: SO-ARM101 Red-Block Pick-and-Place

WALL-OSS-0.5 (X-Square Robot) full parameter fine-tuned on a real SO-ARM101 6-axis arm for the task "grab the red block and place it in the box".

Result

• Real-robot closed-loop: successfully grasps and places (multiple consecutive successes)
• No LoRA, no custom gripper BCE, no execution-side heuristics — pure flow-matching action head
• Validation flow loss 0.012 (vs LoRA baseline plateau 0.017)

Key finding

The previous LoRA fine-tune (0.14% trainable params) failed to close the loop. Full fine-tuning is the canonical path for WALL-OSS adaptation.

Model

• Backbone: Qwen2.5-VL-3B (WALL-OSS-0.5)
• Action space: 26 dims (6 joints + 20 action_padding, single-arm)
• Gripper: continuous flow matching (no binarization)
• Action horizon: 32

Data

• 40 episodes / 6582 frames, 12 fps, dual camera (global + wrist)
• Dataset: https://github.com/wangpengyu-123/so101-round5-dataset

Usage

from safetensors.torch import load_file
sd = load_file("model.safetensors")

See config.yml for the wall-x training config and norm_stats.json for normalization.