cagataydev/smolvla-tictactoe

smolvla_tictactoe_a

Post-tuned SmolVLA policy for the SO-101 arm playing tic-tac-toe by picking and placing coloured blocks.
Trained end-to-end through the strands-robots training abstraction on the public HashtagRobotics/tic-tac-toe-so101-block-a-clean-v1 LeRobot v3 dataset.

This model is the "block-A" post-tune — first of a planned series. Vision + language + 6-DoF joint action, 30 Hz control.

---

TL;DR

Base model	lerobot/smolvla_base
Dataset	HashtagRobotics/tic-tac-toe-so101-block-a-clean-v1 (195 episodes, ~144k frames @ 30fps)
Robot	SO-101 follower (6-DoF, shoulder_pan / shoulder_lift / elbow_flex / wrist_flex / wrist_roll / gripper)
Cameras	2× RGB: observation.images.top (mapped → camera1), observation.images.wrist (mapped → camera2)
Training	20,000 steps · batch 8 · full fine-tune · single H100/Thor GPU
Final loss	0.088 (after 2h 20m)
Trainer	strands_robots.training.create_trainer("lerobot_local", policy_type="smolvla")
Hardware	NVIDIA Thor (aarch64)

---

Quick load

With strands_robots (recommended)

from strands_robots import create_policy

policy = create_policy(
    "lerobot_local",
    pretrained_name_or_path="cagataydev/smolvla_tictactoe_a",
    policy_type="smolvla",
    device="cuda",
    trust_remote_code=True,
)
# → policy.get_actions(observation, instruction=...) returns 6-DoF SO-101 actions.

With raw LeRobot

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("cagataydev/smolvla_tictactoe_a")

In strands_robots simulation (MuJoCo)

from strands_robots import Robot

sim = Robot("so101", mesh=False)
sim.add_camera(name="front", position=[0.50, 0, 0.30], target=[0.22, 0.025, 0.05])
sim.add_camera(name="wrist", position=[0.0, 0.0, 0.05], target=[0.05, 0, 0],
               parent_body="so101/gripper", fov=70.0)

sim.run_policy(
    robot_name="so101",
    policy_provider="lerobot_local",
    policy_config={
        "pretrained_name_or_path": "cagataydev/smolvla_tictactoe_a",
        "policy_type": "smolvla",
        "device": "cuda",
    },
    instruction="pick up the red block and place it on the top-left square",
    n_steps=500,
    control_frequency=30,
)

Camera name contract: SmolVLA base expects observation.images.camera1/camera2.
The dataset uses top + wrist → training was done with a rename map (see below). At inference in sim, mount your two cameras and route them through the same rename_map in policy_config, or name your sim cams top and wrist and let the embodiment do the routing.

---

The task

Post-tuned to control an SO-101 arm to pick up a coloured block and place it on a 3×3 tic-tac-toe grid.
Instructions look like:

• "pick up the red block and place it on the center square"
• "place a blue block on the top-right"
• "put your piece on the middle-left square"

The A variant is trained on the cleaned "block-A" split (single-block-type demonstrations, 195 episodes). Future variants (_b, _multiblock, ...) will follow.

---

Training recipe

Training was driven entirely through the strands_robots.training DX, not raw lerobot.scripts.lerobot_train — this was an end-to-end test of that abstraction on a real public HF dataset.

Full config

from strands_robots.training import TrainSpec, create_trainer

trainer = create_trainer("lerobot_local", policy_type="smolvla")

spec = TrainSpec(
    dataset_repo_id="HashtagRobotics/tic-tac-toe-so101-block-a-clean-v1",
    base_model="lerobot/smolvla_base",
    output_dir="./checkpoints",
    steps=20_000,
    global_batch_size=8,
    method="full",              # full fine-tune (not LoRA, not expert-only)
    save_freq=1000,
    num_gpus=1,
    seed=42,
    streaming=False,            # full local download
    extra={
        "policy_type": "smolvla",
        "job_name": "smolvla_tictactoe_a",
        "rename_map": {
            "observation.images.top": "observation.images.camera1",
            "observation.images.wrist": "observation.images.camera2",
        },
    },
)

result = trainer.train(spec)
# → status=success, checkpoint at result.checkpoint_dir

Hyperparameters

Knob	Value
Steps	20,000
Global batch size	8
Learning rate	LeRobot default cosine schedule (peak ≈ 1.5e-4, end ≈ 2.5e-6)
Optimizer	AdamW
Method	full (all params trainable)
Precision	bf16 mixed
Chunk size	50 (SmolVLA default)
GPUs	1 × NVIDIA Thor
Seed	42
Data	full download, LeRobot v3 format
Save frequency	every 1000 steps
Wall-clock	2h 20m (≈2.4 steps/s average)

Loss curve (from training log)

Step	Samples	Epoch	Loss	Grad-norm	LR
200	2K	0.01	0.352	6.19	1.5e-05
400	3K	0.02	0.254	4.73	4.5e-05
600	5K	0.03	0.273	4.82	7.5e-05
5,000	~40K	0.26	0.147	~2.1	~1.4e-04
10,000	~80K	0.52	0.109	~1.8	~9.8e-05
14,000	~112K	0.72	0.097	~1.7	~4.9e-05
15,000	120K	0.78	0.095	1.71	1.7e-05
18,000	144K	1.00	0.092	1.57	5.1e-06
19,000	152K	1.05	0.087	1.53	3.2e-06
20,000	160K	1.11	0.088	1.61	2.5e-06

Loss dropped from 0.35 → 0.088 over 20K steps. First epoch reached at ~step 18,000 (1× full traversal of 144K frames).

Checkpoints

Automatic checkpoints saved every 1000 steps to checkpoints/{step}/pretrained_model. This upload is the step 20,000 / final checkpoint.

---

Dataset details

Source	HashtagRobotics/tic-tac-toe-so101-block-a-clean-v1
Format	LeRobot v3 (parquet)
Episodes	195
Frames	~144,000
FPS	30
Robot type	so101
State/action dim	6 (joint positions)
Camera resolution	480×640

Feature routing (rename_map):

observation.images.top    → observation.images.camera1
observation.images.wrist  → observation.images.camera2

camera3 is not populated — SmolVLA-base tolerates missing 3rd cam when only 2 of 3 slots are provided, but you should apply the same rename at inference time.

---

Files in this repo

File	Purpose
config.json	Full SmolVLA policy config (frozen at training time)
model.safetensors	906 MB, all trainable params (VLM backbone + action expert)
policy_preprocessor.json + .safetensors	Input normalizers (state/image scaling)
policy_postprocessor.json + .safetensors	Output unnormalizer (action denormalization)
train_config.json	Full training config for reproducibility (steps, batch, lr schedule, seed, …)

policy_preprocessor.json and policy_postprocessor.json are required for correct action denormalization — do not omit them when loading.

---

Reproducibility

To reproduce this exact model on your hardware:

git clone https://github.com/strands-labs/robots
cd robots && pip install -e '.[lerobot,smolvla]'

Then run the shipped training script:

# train_smolvla_tictactoe.py — same script that trained this checkpoint
from strands_robots.training import TrainSpec, create_trainer

trainer = create_trainer("lerobot_local", policy_type="smolvla")
spec = TrainSpec(
    dataset_repo_id="HashtagRobotics/tic-tac-toe-so101-block-a-clean-v1",
    base_model="lerobot/smolvla_base",
    output_dir="./checkpoints",
    steps=20_000,
    global_batch_size=8,
    method="full",
    save_freq=1000,
    seed=42,
    extra={
        "policy_type": "smolvla",
        "rename_map": {
            "observation.images.top":   "observation.images.camera1",
            "observation.images.wrist": "observation.images.camera2",
        },
    },
)
trainer.train(spec)

Wall-clock on NVIDIA Thor: 2h 20m for 20K steps.

---

Evaluation

Post-training in-domain training loss: 0.088.

Real-world evaluation on physical SO-101 hardware and MuJoCo sim (using the strands_robots sim pipeline to recreate the tic-tac-toe scene) is in progress — will publish success rates + video rollouts as a follow-up here + on Telegram/GitHub.

Planned eval metrics:

• Task success rate: block correctly placed on target grid square (10 trials × 9 squares × 3 colours).
• Sim-to-real gap: same policy in MuJoCo scene reconstruction vs. real SO-101.
• Instruction following: text-conditioned generalisation to unseen instructions (colour × position combos).

---

Known limitations

1. Single dataset variant — trained only on the "block-A clean" split. May not generalise to different block colours/shapes without further tuning.
2. Two cameras only — no wrist-only or third-person-only ablation yet.
3. 6-DoF joint control — inherits the SO-101 kinematic constraints (no end-effector cartesian control).
4. Rename map is required at inference — if you name cameras differently, you must apply the same routing.
5. No LoRA / expert-only variant yet — full fine-tune only (906 MB). LoRA variants are planned to drop below 100 MB.

---

Citation & lineage

• Base: lerobot/smolvla_base (SmolVLA architecture, HuggingFace / LeRobot team)
• Dataset: HashtagRobotics/tic-tac-toe-so101-block-a-clean-v1
• Training abstraction: strands_robots.training
• Trainer/uploader: @cagataycali
• Compute: NVIDIA Thor dev kit (aarch64)

Please cite the base model + dataset if you use this checkpoint.

---

Contact

• Issues / questions → strands-labs/robots issues
• Author → @cagataycali
• License → Apache-2.0 (matches base model)