PointmapVLA-RoboCasa

PointmapVLA fine-tuned on the 24 RoboCasa kitchen tasks (MimicGen-300 demonstrations). This is the checkpoint behind the RoboCasa result of the paper. Code, recipes and the evaluation harness: PointmapVLA repository.

What it is

PointmapVLA is a vision-language-action model built on the Qwen3.5-4B vision-language backbone with a pointmap encoder (a copy of the vision tower fed raw metric XYZ pointmaps in the robot-base frame, features added patch-wise to the RGB tokens) and a 32-layer layer-wise flow-matching DiT action expert (4 Euler steps at inference). 6.08 B parameters, bf16.

This checkpoint predicts chunks of 16 RoboCasa actions (15-D) from three camera views (left, right, wrist), the instruction and an 18-D state, from which an 11-D end-effector anchor is fed to the action expert.

Provenance

• Initialised from DAVIAN-Robotics/PointmapVLA-Pretrained (Stage 1 + Stage 2 pretraining on 76.3 M robot frames and 10.7 M 3D spatial VQA samples).
• Fine-tuned for 20,000 optimizer steps, global batch 128 (4 GPUs x 32), AdamW, cosine schedule with 1,000 warm-up steps, learning rates 1e-5 (backbone, pointmap encoder) and 1e-4 (action expert, projections), DeepSpeed ZeRO-2, bf16.
• Data: DAVIAN-Robotics/robocasa-MG_300-r6
  • DAVIAN-Robotics/robocasa-MG_300-pointmap (7,200 episodes, 2.08 M frames, 3 cameras).
• On load from the pretrained checkpoint the four action-head tensors whose shape depends on the layout were remapped (shared end-effector/gripper slots copied, new slots fresh); all other weights transferred unchanged.
• Prompt: Task: {instruction} Mode: <end_effector> Pointmap_origin: <robot> with enable_thinking: false (an explicit empty <think> block).
• Exact config: examples/simBenchmarks/Robocasa_kitchen/train_files/pointmapvla_robocasa_kitchen.yaml in the repository (_80gb.yaml is the same recipe with gradient accumulation).

Files

config.yaml                        architecture and the data recipe (release schema)
dataset_statistics.json            action/state mean/std/min/max/quantiles used for (un)normalisation
base_vlm/                          Qwen3.5-4B config, tokenizer and processor only; no weights
checkpoints/
  model-00001-of-00003.safetensors
  model-00002-of-00003.safetensors
  model-00003-of-00003.safetensors   ~11 GB in total, bf16
  model.safetensors.index.json

State-dict key groups: qwen_vl_interface.model.* (backbone), pointmap_encoder.*, action_model.* (DiT expert), project_layers.* (per-layer projections). The language model head is tied to the token embedding and stored once; the index records it under metadata.tied_weights and the loader re-ties it.

How to load

from starVLA.model.framework.VLM4A.PointmapVLA import PointmapVLA

model = PointmapVLA.from_pretrained("DAVIAN-Robotics/PointmapVLA-RoboCasa")
out = model.predict_action(examples, noise_seed=0)      # {"normalized_actions": (B, 16, 15)}

examples[i] is a dict with image (3 HWC uint8 arrays, 224 x 224), lang, state ((1, 18) float32, normalised), pointmap ((3, 256, 256, 3) float16 metres, robot-base frame) and pointmap_valid ((3, 256, 256) bool). Un-normalise with dataset_statistics.json. In practice, run the policy server from the repository, which does the un-normalisation and the r6 -> axis-angle conversion:

python deployment/model_server/server_policy.py --ckpt DAVIAN-Robotics/PointmapVLA-RoboCasa --port 20000 --use_bf16

Evaluation

Protocol: 24 tasks x 50 episodes, max 720 steps, action chunk 16 executed in full, 4 Euler steps, environment seed 42 (episode i uses 42*100000 + i), three repetitions with policy noise seeds 0, 1, 2; the number is the mean (± sd) over repetitions. One repetition alone typically lands within about ±1.3 points of the mean.

CKPT=DAVIAN-Robotics/PointmapVLA-RoboCasa GPUS=0,1 scripts/eval/run_eval_robocasa.sh

Benchmark	Paper	Our reproduction
RoboCasa kitchen-24, success rate (%)	78.2	77.8 ± 0.6 (3 repetitions)

Intended use and limitations

• Simulation benchmark checkpoint for the RoboCasa kitchen-24 tasks (Panda-Omron, MG-300 distribution). Not validated on physical robots.
• Layouts are fixed: action 15-D (EE position 3 | EE rotation r6 6 | gripper 1 | base motion 4 | control mode 1), horizon 16; state 18-D (base position 3 | base quaternion 4 | EE position 3 | EE rotation r6 6 | gripper 2). For another robot or layout, fine-tune PointmapVLA-Pretrained instead.
• Requires metric depth with known intrinsics and camera-to-world extrinsics for the pointmap input; the model was trained with pointmaps present.

License

MIT (code and weights of this release). Weights are derived from Qwen3.5-4B, distributed under the Apache License 2.0. See NOTICE in the repository.

Citation

TBD. The paper is under double-blind review; the BibTeX entry and author list will be added at camera-ready.