TASX-Command-0.5B

TASX-Command-0.5B is a highly specialized, lightweight language model designed specifically for robotics. It translates natural language (including slang, typos, and complex phrasing) into strict, execution-ready JSON command sequences for ROS2, SLAM, and physical robot control.

By fine-tuning the Qwen2.5-0.5B base model, we created a "robot brain" that is small and fast enough to run locally on edge hardware (like a Raspberry Pi) via llama.cpp while retaining the intelligence to understand complex human intent.

📦 Quantized Versions (GGUF)

For high-performance inference , use these GGUF quants:

TASX-Cmd-0.5B-GGUF (mradermacher) — Includes high-quality iMatrix and IQ quants.
TASX-Cmd-0.5B-Q8_0 (ReXeeD) — Standard high-precision 8-bit quantization.

🌟 Key Features

Strict JSON Output: Never outputs conversational filler; only valid JSON arrays.
Typo & Slang Immunity: Successfully maps messy speech (e.g., "scoot forward lik 3 point 5 meeters") to perfect floats and commands.
Dynamic Location Extraction: Converts any spoken room or location name (e.g., "Professor Xavier's Office") into clean snake_case (e.g., professor_xavier_office).
Physical Constraint Logic: Automatically generates implicit macro sequences (like sit -> stand -> move) for fetching and delivering items without needing explicit user instruction.

🛠️ Supported Actions & Commands

The model is trained to strictly output one or more of the following 20 commands formatted as a JSON array of actions.

1. Teleop (Movement & Speed)

{"type": "teleop", "cmd": "move_forward", "distance": <float>}
{"type": "teleop", "cmd": "move_backward", "distance": <float>}
{"type": "teleop", "cmd": "rotate_left", "angle": <float>}
{"type": "teleop", "cmd": "rotate_right", "angle": <float>}
{"type": "teleop", "cmd": "set_speed", "level": "slow" | "normal" | "fast"}
{"type": "teleop", "cmd": "stop"} (For casual pauses)
{"type": "teleop", "cmd": "e_stop"} (For panicked/emergency stops)

2. Nav2 (Autonomous Navigation)

{"type": "nav2", "cmd": "go_to_waypoint", "target": "<snake_case_string>"}
{"type": "nav2", "cmd": "cancel_goal"}

3. Stunts (Posture & Tricks)

{"type": "stunt", "cmd": "full_sit"}
{"type": "stunt", "cmd": "half_sit"}
{"type": "stunt", "cmd": "stand_up"}
{"type": "stunt", "cmd": "spin", "direction": "clockwise" | "anticlockwise"}

🧠 Advanced Behaviors (Macros)

TASX-Command-0.5B has been taught physical robotics logic. It knows a robot cannot drive while sitting.

If you ask it to perform a delivery (e.g., "Fetch my laptop from the server room and bring it to John's desk"), it will automatically output the required posture macros:

{
  "actions": [
    {"type": "nav2", "cmd": "go_to_waypoint", "target": "server_room"},
    {"type": "stunt", "cmd": "full_sit"},
    {"type": "stunt", "cmd": "stand_up"},
    {"type": "nav2", "cmd": "go_to_waypoint", "target": "john_desk"},
    {"type": "stunt", "cmd": "full_sit"}
  ]
}

TEST SCRIPT

# ============================================================
# ============================================================
from unsloth import FastLanguageModel
import torch

MODEL_PATH = "./tasx_sft_merged"
MAX_SEQ_LENGTH = 512

print("⏳ Loading your fine-tuned TASX model...")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = MODEL_PATH,
    max_seq_length = MAX_SEQ_LENGTH,
    dtype = torch.float16, 
    load_in_4bit = False,  
)

FastLanguageModel.for_inference(model) 

print("\n" + "="*50)
print("TASX ROBOT COMMAND TESTER READY")
print("Type a command in the box below and press Enter.")
print("Type 'quit' or 'q' to stop.")
print("="*50 + "\n")


while True:
    user_text = input("🎤 You: ")
    
    if user_text.lower() in ['quit', 'exit', 'q']:
        print("Stopping inference. Great job!")
        break
        
    if not user_text.strip():
        continue

   
    prompt = f"<|im_start|>user\n{user_text}<|im_end|>\n<|im_start|>assistant\n"
    
    
    inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

    # Generate the output
    outputs = model.generate(
        **inputs,
        max_new_tokens=150,
        use_cache=True,
        temperature=0.1,    
        do_sample=True,
    )

    # Decode the output (slice off the prompt so we only see the assistant's new text)
    response = tokenizer.batch_decode(
        outputs[:, inputs.input_ids.shape[1]:], 
        skip_special_tokens=True
    )[0]
    
    print(f"TASX: {response.strip()}\n")