rskill-topreward_qwen3vl_4b-any-general-nf4

OpenRAL rSkill — TOPReward (Token Probabilities as Hidden Zero-Shot Rewards for Robotics) packaged as an NF4 reward rSkill on top of lerobot 0.6.0's first-party TOPReward. It is a zero-shot reward: it asks an off-the-shelf Qwen3-VL-4B VLM how likely the task instruction is, conditioned on the rollout video, and reads log P("True") back as the signal. Per-frame progress (0–1) comes from a prefix sweep. No actuators. Advisory-only. Apache-2.0 packaging (upstream method MIT; Qwen3-VL-4B-Instruct weights are also Apache-2.0).

Preview

Per-frame progress overlay on LIBERO libero_object episode 0 — task "pick up the orange juice and place it in the basket" — scored live with the NF4 Qwen3-VL-4B backbone (peak 3.13 GB, RTX 4070 Laptop 8 GB):

Start of episode	Mid-reach	Object placed
progress 0.07	progress 0.52	progress 0.86

HF model cards render images but do not embed HTML5 <video>, so the three frames above (start / middle / end) stand in for the clip. The full overlay video is media/progress.mp4 in this repo (143 frames, downloadable). Regenerate everything with tools/topreward_per_frame_demo.py --media-dir media.

Quick Start

ral skill install hf://OpenRAL/rskill-topreward_qwen3vl_4b-any-general-nf4

from openral_core.schemas import RSkillManifest

manifest = RSkillManifest.from_yaml("rskills/topreward-qwen3vl-4b-nf4/rskill.yaml")
assert manifest.kind == "reward"
assert manifest.role == "s2"
assert manifest.reward.progress_range == (0.0, 1.0)

What It Does

Runs parallel to a VLA policy and scores the rollout so the Reasoner can tell whether a skill is making progress or is done — without any hand-written success detector. Given the rollout's RGB frames plus the task instruction it produces a per-frame normalized progress signal in [0, 1], queried on demand by the Reasoner. It never actuates and never gates motors; its output is advisory input to the replanning ladder.

Field	Value
Actions	monitor
Objects	task progress, task success
Scenes	tabletop, kitchen, indoor, manipulation
Embodiment	any (embodiment-agnostic reward monitor)

Why a reward model alongside the VLA

The VLA emits actions but no notion of "am I done / am I stuck". TOPReward fills that gap zero-shot — no reward head to train, no per-task labels. Because it is a frozen general VLM prompted with the instruction, it transfers across embodiments and tasks, at the cost of an uncalibrated (per-episode min-max) scale rather than a physically-calibrated success bar.

How it works — architecture and upstream model

TOPReward's model (lerobot.rewards.topreward.TOPRewardModel) wraps transformers' Qwen3VLForConditionalGeneration. It builds the prompt

<video> The above video shows a robot manipulation trajectory that completes the
following task: <instruction> Decide whether the above statement is True or not.
The answer is: True

label-masks all but the final token, and returns log P("True" | video, instruction) as one clip-level scalar.

Per-frame progress is lerobot's native prefix sweep (lerobot.rewards.topreward.compute_rabc_weights): score growing trajectory prefixes frames[0:k] at a set of anchor lengths, min-max normalise the raw log-probs per episode, then interpolate back to one value per frame — a [0, 1] progress curve. The OpenRAL runtime feeds this the same rolling RGB buffer the co-active VLA uses.

• Backbone: Qwen/Qwen3-VL-4B-Instruct (zero-shot; no fine-tuned weights).
• Quantization: NF4 (bitsandbytes, double-quant, bf16 compute). lerobot's TOPRewardModel loads bf16 with no quant knob, so the OpenRAL backend subclasses it to inject a BitsAndBytesConfig (see tools/topreward_per_frame_demo.py::NF4TOPRewardModel).
• transformers: 5.x works directly — no version downgrade pin.

Runtime

Inference contract

Direction	Key	Shape	Notes
in	rolling RGB window	(T, 3, H, W) uint8/float	same camera stream as the VLA
in	task instruction	str	required (instruction_required: true)
out	per-frame progress	(T,) float in [0, 1]	prefix sweep + per-episode min-max

Each prefix forward is capped to 8 frames (evenly tail-cropped) to bound the Qwen3-VL video activation on 8 GB.

Validated live

Run on RTX 4070 Laptop (8 GB) against real LIBERO libero_object episode 0 (143 frames, a success demo) via tools/topreward_per_frame_demo.py:

Metric	Value
Backbone	Qwen3-VL-4B-Instruct (NF4)
Peak VRAM	3.13 GB
Progress, first 20% of episode	0.41
Progress, last 20% of episode	0.92
Curve	rises 0 → 1 as the juice reaches the basket

Supported robots and embodiments

Embodiment-agnostic (embodiment_tags: ["any"]). A reward monitor scores any rollout video + task instruction, so it is exempt from the rSkill↔robot embodiment gate. Validated on a Franka Panda LIBERO scene; nothing about the model is Franka-specific.

Sensors and Observation Contract

Consumes one RGB camera stream (modality: rgb, ≥ 224×224) — the same frames the co-active VLA observes. No actuators required. The instruction string is supplied by the Reasoner from the active task.

Manifest Summary

Field	Value
name	OpenRAL/rskill-topreward_qwen3vl_4b-any-general-nf4
kind	reward
role	s2
weights_uri	hf://Qwen/Qwen3-VL-4B-Instruct (zero-shot backbone)
quantization	NF4 / bitsandbytes (int4, bf16 compute)
min_vram_gb.int4	3.2 (measured 3.13 peak)
reward.progress_range	[0.0, 1.0]
reward.success_threshold	0.8 (advisory; uncalibrated)
reward.target_fps	2.0
paper_url	https://arxiv.org/abs/2602.19313

See rskill.yaml for the full manifest.

License

rSkill packaging is Apache-2.0 (all OpenRAL code is uniformly Apache-2.0). The TOPReward method is MIT. The wrapped Qwen3-VL-4B-Instruct weights are also Apache-2.0 (Qwen/Qwen3-VL-4B-Instruct, license: apache-2.0), so an NF4-quantized copy is freely redistributable — the whole stack is Apache-2.0 / MIT. Cite the TOPReward paper (Chen et al., 2026).