tiantianx/aloha2-smolvla-banana-apace-d20-scratch-matched

A-PACE ALOHA banana: scratch JEPA and matching adapters, d=20

This release contains the validation-selected JEPA at step 20,160 and two adapters trained against that exact checkpoint. The full JEPA training run completed 25,200 steps; the released checkpoint is selected, not the last step. This is a separate repository from the earlier 12,600-step interim release.

File	Meaning
jepa_training/best.pt	JEPA, 20,160 steps, selected by held-out physical arm-joint MSE
adapter_training/step001500.pt	Optional matching adapter; lowest whole-chunk and first-three-action arm MSE among scanned adapters
adapter_training/step002000.pt	Optional matching adapter; lowest gripper MSE among scanned adapters

Both adapter payloads bind to the JEPA SHA256 in RELEASE.json. Neither is compatible with the previous 12,600-step release or the old JEPA. Pure JEPA remains a necessary baseline: the adapter improvements are small and not established as robust closed-loop gains.

Pinned base and data

• Base policy: yhong96/aloha2_smolvla_banana_v2, revision 558170cf34a8a5c83e5ce68944846186a82a4b80.
• Dataset: yhong96/aloha2_banana_insertion, revision 213b3da1b40e35ab67278eafe2415461982ce65e.
• Native action horizon: 50; delay: 20 control steps, nominally 400 ms at 50 Hz.
• JEPA input: current native condition and 20 committed absolute 14D ALOHA command vectors.
• JEPA output: future visual condition and native normalized joint/gripper state.

Download and use with the existing ALOHA A-PACE rollout package

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="tiantianx/aloha2-smolvla-banana-apace-d20-scratch-matched",
    local_dir="models/apace_d20_matched",
)

The repository is private: authenticate using an account that has access. The base policy and the existing rollout package are required separately.

For pure JEPA, set the policy server option below and leave --aca-checkpoint unset:

--jepa-checkpoint models/apace_d20_matched/jepa_training/best.pt

For the 1,500-step adapter, additionally set:

--aca-checkpoint models/apace_d20_matched/adapter_training/step001500.pt

For the 2,000-step comparison, replace that adapter path with models/apace_d20_matched/adapter_training/step002000.pt. Do not use a different JEPA checkpoint with either adapter.

Use the pinned base model and its native processors (apace_model_binding.json). Supply the commands that actually execute during the delay. At scheduled handover, start the future-conditioned chunk at index 0; do not skip another 20 actions. The existing runtime converts generated joint offsets to absolute targets using the predicted handover state. Grippers remain absolute. Do not apply this conversion twice. This release does not modify controller or scheduling behavior.

Validation

JEPA: 8 held-out episodes, 2,240 windows. The selected checkpoint has:

• 12-arm-joint RMSE: 1.454 degrees (old JEPA: 3.303).
• Left wrist angle RMSE: 1.869 degrees (old: 6.762).
• Left wrist angle maximum error: 6.213 degrees (old: 30.827).
• Visual feature MSE: 1822.034 (old: 1728.156; lower is better).

The state prediction improved, while visual prediction remains worse than the old model.

Adapter: 128 windows from the same 8 held-out episodes, identical sampling noise, full native action sampling, absolute-command reconstruction. Relative MSE reductions against pure new JEPA:

Adapter step	Whole chunk, arm	First 3 actions, arm	Gripper
500	2.21%	0.74%	5.53%
1,000	1.62%	0.70%	5.50%
1,500	2.63%	1.00%	6.28%
2,000	2.10%	0.52%	7.26%

The paired episode-bootstrap 95% intervals include zero. There is no consistent arm-error improvement with increasing steps from 500 to 2,000. Checkpoint selection is exploratory on this validation set, not an independent test result. These metrics are not physical jerk; robot success and handover smoothness remain unverified for this release.

Training details

JEPA was initialized from scratch, using cached frozen-policy features. Effective batch 16, AdamW peak LR 2e-4, OneCycle, 20 epochs. Training balances visual MSE, physical arm-joint MSE and gripper MSE by train-only scales, plus cosine loss. The run was recovered at step 12,000 after interruption, with optimizer/scheduler restored. Its earlier save omitted RNG state, so recovery is not bitwise identical to uninterrupted training.

Adapters use a zero-initialized residual, frozen JEPA and frozen VLA, 2,000 updates, effective batch 4 and LR 1e-4. Labels subtract the predicted handover arm state from absolute demonstration commands; grippers remain absolute. This aligns training targets with runtime action reconstruction. The adapter adjusts visual condition and does not itself correct the JEPA state prediction. No physical jerk or boundary loss is included.

All three weights passed loading through the existing rollout runtime classes; adapter/JEPA identity was checked locally. SHA256SUMS.json provides file hashes. Raw validation records and episode-level comparisons are included under validation/.