CARE-Drive: Coupled Alternating Refinement for Ego-trajectories

Final inference checkpoints for CARE-Drive, accompanying the training and inference code.

Standalone checkpoint loading requires CARE-Drive 0.1.1 or later (compatible code revision).

CARE-Drive combines a predictive Qwen3-VL backbone with a waypoint residual expert and a position-conditioned heading expert. The position expert corrects eight future waypoints with bounded metric residuals; the heading expert reconstructs heading from eight adjacent angular increments conditioned on the corrected path.

Released weights

Component	Stage	File
Predictive backbone	A0-DYNAMIC-L1, epoch 5	A0-DYNAMIC-L1/ — checkpoint + assets
Position expert	C3-XY-EP1, epoch 5	C3-XY-EP1/epoch_005.pt
Final heading expert	D1-H-ADAPT, epoch 5	D1-H-ADAPT/epoch_005.pt

These are the final model tensors used for the paper's 88.837 PDMS system on 12,146 NAVSIM v1 navtest scenes. The score uses the original official CPU evaluation. The release extracts model weights from the accepted training checkpoints, preserving every tensor's value, name, shape, and dtype. Optimizer, scheduler, RNG, and internal experiment metadata are omitted. This is an inference/model-initialization release rather than an optimizer-resume snapshot.

checkpoint_manifest.json records the exported file sizes and SHA256 checksums, original checkpoint checksums, and matching tensor digests. Backbone tensors retain their original BF16/FP32 dtypes; both experts retain FP32. The two experts have 78,420,113 and 78,418,577 parameters, respectively. Companion backbone assets add approximately 16 MB and are included by the download command below.

Download and run

Install the CARE-Drive environment, then run from the code repository root:

hf download maxiao123/CARE-Drive \
  --include "A0-DYNAMIC-L1/*" "C3-XY-EP1/*" "D1-H-ADAPT/*" "checkpoint_manifest.json" \
  --local-dir runs/care-drive

The backbone includes A0-DYNAMIC-L1/assets/: its final architecture, tokenizer with CARE-Drive query tokens, chat template, and image/video processors. The loader constructs the model locally from this bundle and directly restores the trained checkpoint. No original Qwen weights or V-JEPA model are needed for inference. Keep the assets/ directory beside the backbone checkpoint when moving files. Use CARE-Drive code version 0.1.1 or later; rerun the download command to add the assets to an existing download.

export CARE_DRIVE_DATA=/absolute/path/to/data

Set your device in configs/inference.yaml; pass the observation manifest and sensor directory to the inference command. Prepare an observation-only manifest using the data guide, then use the inference command:

export CARE_DRIVE_PERFORMANCE_PREFLIGHT=/absolute/path/to/performance/inference.json
python -m care_drive.infer \
  --config configs/inference.yaml \
  --manifest "$CARE_DRIVE_DATA/manifests/navtest.jsonl" \
  --sensor-root "$CARE_DRIVE_DATA/sensor_blobs/test" \
  --backbone runs/care-drive/A0-DYNAMIC-L1/epoch_005.pt \
  --position runs/care-drive/C3-XY-EP1/epoch_005.pt \
  --heading runs/care-drive/D1-H-ADAPT/epoch_005.pt \
  --batch-size 1 --output runs/care-drive/navtest_predictions.jsonl

Formal batch execution uses the codebase's performance setup. V-JEPA2.1 supplies training targets and is not needed for inference.

Each input contains four historical frames from three cameras, ego pose history, velocity, acceleration, and a route command. Each output is eight (x, y, heading) poses at 0.5-second intervals, covering four seconds. Coordinates use the current ego rear-axle frame, with meters for position and radians for heading.

Provenance and validation

The backbone was trained on 92,772 scenes; the expert chain used 92,614 scoreable scenes. The released configuration records the paper's quantization and architecture. Exact cohort manifests are not included.

Export verification compares the complete tensor content against the original accepted checkpoints. All three checkpoints passed strict loading with the published CARE-Drive implementation, including the backbone tokenizer contract. Loaded tensor values and dtypes match the exports exactly, and a single synthetic feature sample passed coupled expert generation. Standalone loading also passed with an empty Hugging Face cache, networking disabled, and original Qwen weight loading forbidden. Backbone parameters, rotary buffers, deterministic forward outputs, and multimodal preprocessing matched the previous loader exactly. A complete raw-observation CLI run on one synthetic scene also produced a finite 8-by-3 trajectory under the same offline restrictions. The paper's benchmark result comes from the original evaluation; the newly packaged entry points have not been used to rerun full navtest. See the code repository's validation record for the checked scope.

The model is intended for research in driving trajectory planning. Data and maps are obtained separately. Original Qwen weights and the V-JEPA teacher are required only for training from scratch. This repository distributes CARE-Drive's trained weights under Apache-2.0; upstream dependencies retain their respective licenses.