Read this first. This is a pipeline smoke-test artifact, not a usable policy. It was trained on 7 episodes (1253 frames) for 10,000 steps at global batch 64 — roughly 510 epochs over the same handful of trajectories. It is heavily overfit by construction. It was produced to verify that training and inference run end to end on this robot's data, and it should not be taken as an indication of what RLDX-1 can do.

What it does

Predicts 16-step absolute joint-target chunks for both arms and both hands from a single head camera and the current joint state.

Modality	Keys	Dims
video	`cam_head`	4 frames, stride 2 (`delta_indices=[-6,-4,-2,0]`)
state	`left_arm`, `right_arm`, `left_hand`, `right_hand`	7 / 7 / 20 / 20
action	same 4 groups, absolute joint targets	horizon 16
language	`annotation.human.task_description`	—

The robot's raw observation.state / action vectors are 57-dim; this model uses 54 of them. head_joint1/2 and lift_joint are excluded because they never move anywhere in the training data.

Actions are absolute joint targets (ActionRepresentation.ABSOLUTE, ActionType.NON_EEF), not deltas and not end-effector poses.

Known limitations

Overfit. ~510 epochs over 7 episodes. Expect it to reproduce the training trajectories and generalize to nothing.
The language conditioning is meaningless. The training dataset's tasks.jsonl contains the placeholder string "1611" — a task ID, not an instruction — so the model was conditioned on that literal string for every sample. It has no instruction-following ability.
Head camera only. Wrist cameras exist in the source data but were not used.
Hands are weakly supervised. Only 2 of the 7 episodes move the fingers at all; the rest move the arms with the hands held fixed.

Evaluation

Open-loop, against ground-truth actions on held-in trajectories (there is no held-out split — the dataset is too small):

checkpoint	traj 1 MSE / MAE	traj 2 MSE / MAE
20 steps (smoke)	0.04499 / 0.14648	0.04576 / 0.14857
10,000 steps (this)	0.00718 / 0.03328	0.00727 / 0.03133

MAE 0.033 rad ≈ 1.9°. These are training trajectories, so the numbers measure fit, not generalization. Note also that the flow-matching sampler is not seeded: repeated evaluations of an unchanged checkpoint vary by roughly ±15%.

Peak inference VRAM at batch 1: 15.83 GiB (bf16, 6.9B parameters).

Usage

Requires the RLDX-1 codebase.

bash

python rldx/eval/open_loop_eval.py \
    --model-path learner1119/RLDX-1-FT-ffw-sh5 \
    --dataset-path /path/to/ffw_sh5_0812 \
    --embodiment-tag GENERAL_EMBODIMENT \
    --traj-ids 1 2 --action-horizon 16 --steps 64

Or serve it:

python rldx/eval/run_rldx_server.py --model-path learner1119/RLDX-1-FT-ffw-sh5

Observation keys follow {modality}.{key}: video.cam_head, state.left_arm, state.right_arm, state.left_hand, state.right_hand, annotation.human.task_description.

No --modality-config-path is needed — the modality config is baked into the processor saved alongside these weights.

Training


Base	`RLWRLD/RLDX-1-PT` (video-input, 6.9B)
Data	`learner1119/ffw_sh5_0812`, LeRobot v2.1, 7 episodes / 1253 frames @ 10 fps
Embodiment tag	`GENERAL_EMBODIMENT`
Steps / batch	10,000 / 64 global
LR / schedule	1e-4, cosine, 5% warmup
Hardware	4× NVIDIA B200

Only the inference artifacts are published here (weights, processor, experiment config). The DeepSpeed optimizer state is not included, so this checkpoint can be used for inference or as a fine-tuning base, but not to resume the original run.

License

Licensed under the RLWRLD Model License v1.0 — see LICENSE.md for the full text.

This is a Derivative Model of RLWRLD/RLDX-1-PT. The license is non-commercial, requires attribution, and is share-alike: if you distribute a model derived from this one, it must carry these same terms. Section 3.5 additionally prohibits military, weapons, and non-consensual surveillance applications.