Cosmos3 Edge Policy DROID GenW8A8

Base model | Cosmos Lite | Benchmark | Cosmos 3 report

Run NVIDIA's 4B Cosmos3 Edge DROID robot policy on a single 24GB RTX 4090. This self-contained GenW8A8 deployment is the recommended Cosmos Lite tradeoff for Edge.

At A Glance

Model	Quant.	VRAM (GB)	Request p50 (ms)	RoboLab-120 SR (%)
Cosmos3 Edge 4B	GenW8A8	8.79	331.4	19.25

The release protocol is one RTX 4090, batch size one, guidance 3, two UniPC denoise steps, shift 5, and a 32-action chunk. The result is 231/1,200 closed-loop rollouts across 120 RoboLab tasks, 10 episodes per task, using default instructions. Request latency excludes simulator time.

A representative Edge GenW8A8 closed-loop RoboLab timing trace is shown above. It uses the validated guidance 3 / two-step sampler.

Model Overview

Cosmos3 Edge is NVIDIA's 4B Mixture-of-Transformers world action model for DROID. It predicts robot actions from language, RGB observations, and robot state. Cosmos Lite changes the inference representation and runtime; it does not retrain or distill the policy.

GenW8A8 applies dynamic per-token FP8 E4M3 W8A8 to 168 action-generation linear modules. The other 168 selected language/MoT linear modules use packed W4A16; activations outside the W8A8 modules remain BF16. The bundle is self-contained and does not require the BF16 source checkpoint or calibration dataset at deployment time.

Calibration And Evaluation

No additional policy training is performed. W4 scales use input-channel statistics from 128 successful episodes in the official nvidia/Cosmos3-DROID training split. Calibration does not update model parameters and uses no RoboLab evaluation episode. FP8 activation scales are computed dynamically per token at inference time.

The public result above covers all 120 RoboLab tasks. Full protocol, confidence intervals, topology, and memory definitions are in the Cosmos Lite benchmark.

Policy Interface

The validated RoboLab client composes three camera views into a 640x540 RGB input and maps it to the model's 736x544 inference bucket. It returns a 32x8 DROID joint-position action chunk; the standard integration executes all 32 actions before requesting another chunk.

Model precision is fixed by this artifact. Guidance, denoise steps, and shift are independent runtime controls. The release config defaults to guidance 3, two denoise steps, and shift 5.

Benchmark Results

The primary release result is the full RoboLab-120 evaluation above. It is compared with Edge BF16 in the main benchmark; sampler and legacy quantization experiments are separated into RoboLab ablations. The recommended role of this artifact is low-latency Edge inference; the official NVIDIA BF16 checkpoint remains the original-weight alternative.

Artifact	VRAM (GB)	Request p50 (ms)	RoboLab-120 SR (%)
Edge BF16	9.20	582.0	20.92
Edge GenW8A8	8.79	331.4	19.25

Quickstart

Install the Hugging Face CLI, clone the runtime, and download this self-contained bundle:

python -m pip install -U "huggingface_hub[cli]"
hf download "XXXXyu/Cosmos3-Edge-Policy-DROID-GenW8A8" --local-dir /data/cosmos_lite/edge_genw8a8

git clone https://github.com/xxxxyu/cosmos-lite.git
cd cosmos-lite
CUDA_VISIBLE_DEVICES=0 examples/robolab_quant/pipeline.sh setup --with-sage

Validate the bundle before loading it:

BUNDLE_DIR=/data/cosmos_lite/edge_genw8a8 \
STRATEGY=gen_branch_w8a8 \
examples/robolab_quant/pipeline.sh validate

For serving and closed-loop RoboLab evaluation, follow the RoboLab pipeline guide.

Quantization And Bundle Details

Field	Value
Strategy	`gen_branch_w8a8`
W4 / W8A8 generation modules	252 / 252 (Nano); 168 / 168 (Edge)
Activation quantization	Dynamic FP8 only in the generation branch; other activations remain BF16
Calibration	W4 scales from 128 successful DROID training episodes; no RoboLab evaluation data

Format Note

This is a Cosmos Lite deployment bundle, not a generic GPTQ or AWQ checkpoint. The packed weights, precision map, residual tensors, processor assets, and runtime manifest must be loaded through Cosmos Lite. Weight-only artifacts keep activations in BF16; GenW8A8 artifacts use FP8 W8A8 only in the generation branch and retain the documented BF16 paths elsewhere.

The default server binds to 127.0.0.1:8000 and has no built-in TLS or authentication. Do not expose it directly to an untrusted network.

Provenance

Base policy: nvidia/Cosmos3-Edge-Policy-DROID
Runtime: Cosmos Lite
Bundle contents include runtime configuration, source revisions, file sizes, and SHA256 hashes.

Limitations And Safety

Simulation results do not establish transfer to a new robot, camera layout, task distribution, or action contract. Validate those changes with paired closed-loop rollouts. This policy is not real-robot safety certified. Real deployment requires an independent E-stop, watchdog, motion limits, stale-command rejection, and operator supervision.

This community quantization is not an NVIDIA product.