forcevla-flexiv-tactar-june26-quat-effort

ForceVLA Flexiv Tactar June26 quaternion+effort checkpoint - step 19999

ForceVLA / pi0-guidance checkpoint fine-tuned on the real_single_flexiv_tactar_june26 LeRobot dataset.

• Training config: forcevla_lora_real_single_flexiv_tactar_june26_quat_effort
• Checkpoint step: 19999 from a 20 000-step run
• Base weights: OpenPI pi0_base
• Images: external camera + left wrist + right wrist as configured in ForceVLA
• State representation: xyz + quaternion(wxyz) + gripper + 6D effort, padded to 32 dims
• Action representation: delta xyz + relative rotation vector + gripper, padded to 32 dims
• Important flags: use_full_state=True, quat_rotvec_actions=True, effort_key=observation.effort
• Assets: assets/real_single_flexiv_tactar_june26/norm_stats.json
• Upload format: inference-only Orbax JAX checkpoint: params/ + assets/ + _CHECKPOINT_METADATA

train_state/ is intentionally not uploaded because it contains optimizer state and is not needed for inference.

Download

huggingface-cli download magic0/forcevla-flexiv-tactar-june26-quat-effort --local-dir ./checkpoints/forcevla_quat_effort_19999

Load in ForceVLA/OpenPI

Point checkpoint_dir at the directory containing params/ and assets/. Do not point it directly at params/.

from openpi.training import config as _config
from openpi.policies import policy_config

train_config = _config.get_config("forcevla_lora_real_single_flexiv_tactar_june26_quat_effort")
policy = policy_config.create_trained_policy(
    train_config,
    "./checkpoints/forcevla_quat_effort_19999",
)

If your inference server exposes an explicit asset id argument, use:

--asset_id real_single_flexiv_tactar_june26

Internally, loading uses:

• weights: openpi.models.model.restore_params(checkpoint_dir / "params", dtype=jnp.bfloat16)
• normalization: checkpoint_dir / "assets" / "real_single_flexiv_tactar_june26" / "norm_stats.json"