OpenRAL/rskill-rldx1_ft-franka_panda-libero_spatial-nf4

rskill-rldx1ft-frankapanda-libero_spatial-nf4

openral rSkill — RLWRLD's RLDX-1-FT-LIBERO (Qwen3-VL-8B backbone + Multi-Stream Action Transformer, ~6.9 B params), packaged for the openral robot agent framework.

This package wraps the upstream RLDX-1 LIBERO finetune with a rskill.yaml manifest that adds capability checking, license surfacing, latency budgets, and openral registry integration. It does not copy model weights. Weights are downloaded directly from the upstream HF repo at first use.

License — read this first

RLDX-1 weights ship under the RLWRLD Model License v1.0 — a non-commercial, attribution-required license that explicitly forbids military / weapons / non-consensual-surveillance use.

• —Loader posture: RSkillManifest.is_commercial_use_allowed → False.
• —Activation: set OPENRAL_ALLOW_NONCOMMERCIAL=1 for research use.
• —Commercial deployment: requires a separate vendor agreement with RLWRLD.

The wrapping openral code is Apache-2.0; the wire-protocol client is a re-implementation against the upstream Apache-2.0 server source (rldx/policy/server_client.py).

Architecture (one paragraph)

RLDX-1 is a vision-language-action model that pairs a Qwen3-VL-8B-Instruct visual backbone (frozen tokenizer; 64 "cognition tokens" condensed via a perceiver) with a Multi-Stream Action Transformer (MSAT) — an MM-DiT extension running three coupled streams (cognition / physics / action) and a flow-matching action head producing a 16-step action chunk per forward. Real-Time Chunking (RTC) is supported but disabled in this manifest's defaults. See the upstream repo at https://github.com/RLWRLD/RLDX-1 for the model class.

Supported robots

The RLDX-1 family ships sibling rSkills for SIMPLER (Google / WidowX), RoboCasa Kitchen / RC365, GR-1, DROID, and the PT foundation checkpoint — all delegate their architecture / license / quantization sections to this README via the openral:rskill-readme-delegates-to: marker.

Sensors required

Two RGB streams + an 8-D proprio state. The sidecar's observation adapter (openral_sim.policies.rldx) stacks the last four frames per camera to satisfy the GENERAL_EMBODIMENT 4-step video horizon, so callers only need to publish the current frame per tick.

Manifest summary

Full schema: `openral_core.schemas.RSkillManifest`.

Auto-managed sidecar (why this rskill is unusual)

RLDX-1 cannot run in-process with openral. We evaluated the in-process path and found it impractical:

1. 1.The rldx Python package pins requires-python = "==3.10.*" plus strict majors on numpy==1.26.4 / torch==2.7.0 / transformers==4.57.0 / flash-attn==2.7.4. Our workspace is Python 3.12 with numpy>=2 / torch>=2.10 / transformers>=5 (CLAUDE.md §3); downgrading would break smolvla, pi05, xVLA, ACT, DP.
2. 2.config.json declares architectures = ["RLDX"] — a custom class that exists only in the rldx package. The HF checkpoint does not ship modeling_rldx.py, so there is no trust_remote_code escape from AutoModel.from_pretrained requiring the local rldx package.
3. 3.Force-installing rldx with --no-deps cascades through 15+ packages with major-version-incompatible APIs (albumentations 2.x vs 1.4, lmdb, av, dm-tree, …) — and even past the imports the model load runs through transformers 5.x against rldx code written for 4.57.
4. 4.Reimplementing the policy in openral would mean porting ~25 kLOC of custom Triton kernels + MSAT flow-matching code — out of scope.

So we run RLDX-1 as an out-of-process sidecar, communicating over its native ZMQ + msgpack wire protocol (aligned with the out-of-process sidecar wire-protocol convention). The openral_sim.policies.rldx adapter auto-manages the sidecar lifecycle so end users never invoke the boot helper by hand: it pings the port, forks tools/rldx_sidecar.py if no server is up, polls until the server answers, and tears the child down on close(). The single 3.10 venv at ~/.cache/openral/rldx-sidecar/source/.venv is reused across every RLDX-1 rSkill — one cached env on disk, not one per checkpoint.

                     ┌──────────────────────────────────────┐
 openral sim run …       │  openral main venv (py 3.12)         │
        │            │   ↳ openral_sim.policies.rldx        │
        ▼            │       (ZMQ REQ + msgpack client +    │
 SimRunner ─────────►│        auto-spawn lifecycle)         │
        ▲            └──────────────┬───────────────────────┘
        │                           │  forks once, then
        │                           │  tcp://127.0.0.1:5555
        │                           ▼
        │            ┌──────────────────────────────────────┐
        │            │  rldx sidecar venv (py 3.10,         │
        │            │  ~/.cache/openral/rldx-sidecar/...)  │
        └────────────│   ↳ python -m rldx.eval.run_rldx_server
                     │       --use-sim-policy-wrapper       │
                     │       (Qwen3-VL backbone + MSAT)     │
                     └──────────────────────────────────────┘

End-user workflow — one command:

openral sim run --config scenes/benchmark/libero_spatial.yaml --rskill rskills/rldx1-ft-libero-nf4 \
            --rskill rskills/rldx1-ft-libero-nf4

The first run on a fresh host clones https://github.com/RLWRLD/RLDX-1, builds the 3.10 venv with uv sync, downloads the ~14 GiB bf16 checkpoint, and NF4-quantizes the backbone. Subsequent runs reuse all of that.

Manual / debugging boot: if you want to keep a long-running sidecar (e.g. on a shared GPU host) or debug a server crash, set OPENRAL_RLDX_AUTO_SPAWN=0 and run the helper yourself:

python tools/rldx_sidecar.py \
    --model RLWRLD/RLDX-1-FT-LIBERO \
    --port 5555 \
    --quantization nf4
OPENRAL_RLDX_AUTO_SPAWN=0 OPENRAL_RLDX_PORT=5555 \
    openral sim run --config scenes/benchmark/libero_spatial.yaml --rskill rskills/rldx1-ft-libero-nf4 \
                --rskill rskills/rldx1-ft-libero-nf4

Quantization — 8 GiB GPU?

Stock weights are bf16, ~14 GiB on disk. The sidecar --quantization nf4 flag NF4-quantizes the Qwen3-VL backbone via bitsandbytes (weight-only, compute_dtype=bf16) and leaves the MSAT diffusion head at bf16 — quantizing the flow-matching head silently corrupts action quality. Approximate VRAM:

Measured on the bundled RTX 4070-mobile 8 GiB:

• —Stock bf16 path OOMs during state-dict load (~7.45 GiB allocated of 7.62 GiB usable, before the MSAT head even allocates).
• —`--quantization nf4` works and fits — VRAM levels off at 6.3 GiB after weight load on the 8 GiB card. The launcher applies NF4 via a transformers.AutoModel.from_pretrained monkey-patch (see tools/rldx_sidecar.py). The upstream policy_loader.py:172 loads the whole stack through a single AutoModel.from_pretrained call, so the monkey-patch quantizes the MSAT diffusion head together with the Qwen3-VL backbone — a known concern for action quality that we accept as the price of fitting on 8 GiB (pass --quantization none on a ≥12 GiB card for the clean reference path).
• —The sidecar ships, listens on tcp://127.0.0.1:5555, answers ping over ZMQ+msgpack, and the openral rldx adapter connects end-to-end.

Open: end-to-end inference round-trip on FT-LIBERO. The shipped FT-LIBERO checkpoint's bundled processorconfig.json registers `generalembodiment (4-frame video horizon [-6,-4,-2,0], OXE-style state keys: eefposabsolute / eefrotabsolute / gripperclose`), not `liberopanda. The openral adapter has been updated to stack 4 historical frames per camera, but the upstream policyloader.py` still rejects `EmbodimentTag.LIBEROPANDA against this checkpoint's processor (KeyError: 'liberopanda'`) — so we boot with `--embodiment-tag GENERALEMBODIMENT and rely on RLDXSimPolicyWrapper's LIBERO↔OXE key alias path (rldx/policy/rldx_policy.py:300-317 + 525-538). Closing the loop on a openral sim run rollout requires one more pass to confirm the action-key reverse-map works and the chunk replays cleanly; see eval/libero_10.json source.status: deferred` for the reproduction CLI.

Gripper convention (GH-133). The policy emits action.gripper in the RLDS dataset convention ([0, 1], 0=close, 1=open); LIBERO/robosuite's OSC controller consumes [-1, +1] with -1=open / +1=close. The openral rldx adapter mirrors the upstream rldx/eval/sim/LIBERO/libero_env.py::{normalize,invert}_gripper_action pair inside _assemble_libero_chunk (helper _rldx_gripper_to_libero) so the gripper actuates correctly. Before that fix, every command landed at ~0 and the Franka gripper never opened or closed.

Recommended GPU: ≥12 GiB VRAM for production runs; 8 GiB is the lower bound the scaffold was validated against.

Quick start

import os
os.environ["OPENRAL_ALLOW_NONCOMMERCIAL"] = "1"

from openral_rskill.loader import rSkill
pkg = rSkill.from_pretrained("OpenRAL/rskill-rldx1_ft-franka_panda-libero_spatial-nf4")

# Or offline:
pkg = rSkill.from_yaml("rskills/rldx1-ft-libero-nf4/rskill.yaml")
print(pkg.manifest.weights_uri)  # → hf://RLWRLD/RLDX-1-FT-LIBERO
print(pkg.manifest.is_commercial_use_allowed)  # → False

Eval reproduction

The eval/*.json blocks in this package are scaffolded with reproduced_locally: false and status: deferred. To close the loop on a ≥ 12 GiB GPU — the adapter auto-spawns the sidecar, so it's a single command:

openral benchmark run \
    --suite libero_10 \
    --rskill rskills/rldx1-ft-libero-nf4

openral benchmark run writes a fresh eval/libero_10.json with reproduced_locally: true and the measured per-task success rates.

Provenance

• —Upstream weights: RLWRLD/RLDX-1-FT-LIBERO
• —Base model: RLWRLD/RLDX-1-PT
• —Visual backbone: Qwen3-VL-8B-Instruct
• —Source code: https://github.com/RLWRLD/RLDX-1 (Apache-2.0)
• —Collection: https://huggingface.co/collections/RLWRLD/rldx-1