aloha2-smolvla-banana-apace

ALOHA banana insertion: A-PACE JEPA + adapter

Paired future-condition predictor (FCP / JEPA) and action-conditioned adapter (ACA) for the frozen SmolVLA policy yhong96/aloha2_smolvla_banana_v2. These are auxiliary checkpoints, not a standalone VLA or a standard PreTrainedPolicy.from_pretrained model.

Files

• jepa_training/best.pt: selected JEPA checkpoint from a completed 25,920-update run.
• adapter_training/final.pt: adapter trained for 2,000 updates with that JEPA frozen.
• */CONFIG.json: training settings; */DONE.json: completion records.
• SHA256SUMS.json: exact checkpoint hashes. The adapter requires the matching JEPA.
• *_offline_validation.json: offline native-interface and full-sampling checks.

The uploaded JEPA is the selected best checkpoint, not necessarily the last training update. Supported deployment delays are 4, 6, 8, 10, 12 control steps at 50 Hz. This release does not support delay 20.

Base policy and data

• Base policy revision: 558170cf34a8a5c83e5ce68944846186a82a4b80.
• Dataset revision: 213b3da1b40e35ab67278eafe2415461982ce65e.
• Dataset: 80 episodes, 72 training and 8 held-out validation episodes.
• Cameras: cam_high, cam_low; state and command dimensions: 14.
• Gripper indices: 6 and 13; native action horizon: 50.
• JEPA training: 20 epochs, effective batch 16, AdamW 2e-4; delay 10 receives half the samples and each other supported delay receives one eighth.

Download

With permission to access this private repository and an authenticated HF client:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="tiantianx/aloha2-smolvla-banana-apace",
    local_dir="aloha2-smolvla-banana-apace",
)

Use the supplied aloha_apace_rollout evaluation integration. In its directory, add these options to scripts/serve_aloha_policy.py:

--backend smolvla \
--checkpoint yhong96/aloha2_smolvla_banana_v2 \
--revision 558170cf34a8a5c83e5ce68944846186a82a4b80 \
--jepa-checkpoint /path/to/aloha2-smolvla-banana-apace/jepa_training/best.pt \
--aca-checkpoint /path/to/aloha2-smolvla-banana-apace/adapter_training/final.pt

Omit --aca-checkpoint for pure JEPA. The corresponding evaluation mode is --baseline jepa_aca or --baseline jepa. Base VLA weights, dataset, robot configuration, and the evaluation code package are not included in this repository.

Action and timing conventions

JEPA receives the absolute 14-dimensional commands committed to execute between observation and handover. It predicts native visual features and normalized future state. The adapter adjusts visual features, leaving the predicted state unchanged. Convert native arm delta outputs to absolute commands using the predicted handover state; the two gripper channels remain absolute. The first newly generated action corresponds to the predicted handover time.

The current integration waits for the scheduled handover if inference finishes early, and stops on a missed deadline. It has not yet implemented late handover by continuing the old chunk. Command steps and wall-clock time must be logged separately on hardware.

Validation scope

Both pure JEPA and JEPA+ACA passed offline full-sampler interface checks for all five supported delays. No real-robot closed-loop success rate or trajectory-quality claim is made. Shared-GPU JEPA+ACA inference measured about 256--267 ms, excluding networking and robot control; this exceeds the 240 ms budget at delay 12. Measurements from separate runs are not a controlled estimate of adapter overhead.

The frozen base VLA remains subject to its upstream terms. No new license grant for upstream assets is made here.