minivla-myCobot280pi-task2-v6

MiniVLA — MyCobot 280 Pi Task2 (v6, Epoch 2)

Vision-Language-Action model fine-tuned for: "Put the yellow block on the shelf onto the black table on the right".

Performance

Model

• Base: prism-qwen25-extra-dinosiglip-224px+0_5b (LIBERO-90 pretrained)
• LLM: Qwen2.5 0.5B, fully unfrozen
• Vision: DINOv2 + SigLIP 224px, fully unfrozen
• Action tokenizer: extra_action_tokenizer (vocab=151921)
• Total params: 1.25B
• Training: 90 trajectories, 2 epochs, lr=2e-5 cosine, batch=1

Quick Start

import os
os.environ["PRISMATIC_DATA_ROOT"] = "./"
os.environ["TOKENIZERS_PARALLELISM"] = "false"

from prismatic.models import load_vla
from deploy.mybot_deploy import MyBotDeploy

# 1. Load model
vla = load_vla("purplehihi/minivla-myCobot280pi-task2-v6", hf_token="dummy")

# 2. Load deployment controller
deploy = MyBotDeploy(stats_path="dataset_statistics.json")
deploy.set_current_encoders(current_encoders)  # from robot

# 3. Per-frame inference
action = vla.predict_action(image, instruction, unnorm_key="mycobot_task2")
encoders, gripper_pwm, info = deploy.step(action)
robot.send_angles(encoders.tolist(), speed=50)
robot.set_gripper(gripper_pwm)

Deployment Pipeline (mybot_deploy.py)

VLA output [dx,dy,dz,droll,dpitch,dyaw,gripper]
  → unnormalize (BOUNDS: q01/q99)
  → current EEF + delta = target EEF
  → numerical IK (damped least squares)
  → 6 joint angles → motor encoders
  → gripper [0,1] → PWM [0,100]

The IK uses damped least squares (Levenberg-Marquardt) with numerical Jacobian. Accuracy: ~0.3mm position, ~0.2° orientation (100 random tests).

Normalization Bounds

Files

• model.pt — Fine-tuned weights (step 60920, loss=2.7e-5)
• dataset_statistics.json — Normalization stats with BOUNDS q01/q99
• config.json — Training config and metadata
• mybot_deploy.py — Edge deployment: VLA output → encoder commands + IK