#本项目以Gr00t轨迹预测模型为例，通过 HTTP Trajectory Server 的方式接入 InternNav 的 Habitat VLN 评测流程 #便于后期更换模型接入接口等 主函数：InternNav/scripts/eval/eval_main.py 用于启动整个推理 用于推理：InternNav/scripts/eval/server_Gr00t.py 完成构造输入，调用模型启动推理，返回结果action。更换模型时照着这个文件的内容仿写一个，换为自己的逻辑即可 用于“HTTP 插头”：InternNav/internnav/evaluator/HTTPTrajectoryClient.py 此文件只用于继承一个BaseTrajectoryClient类，是HTTP Trajectory Server 的 Client，更换模型时按照模型需要使用正确方式包裹发送即可 基本类即函数定义：InternNav/internnav/evaluator/final_habitat_vln_evaluator.py 更换模型后如果有新的参数或逻辑，可以在这里添补

环境配置：

1.创建环境

conda create -n python=3.10 libxcb=1.14 conda activate

2.安装pytorch

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1
--index-url https://download.pytorch.org/whl/cu118

3.安装依赖

pip install -e .[model] --no-build-isolation

4.安装habitat

conda install habitat-sim==0.2.4 withbullet headless -c conda-forge -c aihabitat git clone --branch v0.2.4 https://github.com/facebookresearch/habitat-lab.git cd habitat-lab pip install -e habitat-lab # install habitat_lab pip install -e habitat-baselines # install habitat_baselines

pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu124 cd Path/to/InternNav/ pip install -e .[habitat]

运行方法：

1.首先开启目标模型服务

uvicorn InternNav.scripts.eval.server_Gr00t:app
--host 127.0.0.1
--port 9000

2.接着运行eval文件

python scripts/eval/eval_main.py --model_path /data/sjh/GR00T-Internva/output_uav/checkpoint-300000 --continuous_traj --output_path result/Gr00t/val_unseen_32traj_8steps --save_video

2026.4.18更新：

增加r2r和rxr评测脚本及配置文件，修改了之前的bug 如果要运行sub_ep： r2r:python scripts/eval/eval_main.py --habitat_config_path scripts/eval/configs/our_benchmark_config_sub_r2r.yaml --gr00t_port 9000 --output_path /data/sjh/InternNav/output --save_video --num_history 12

rxr:python scripts/eval/eval_main.py --habitat_config_path scripts/eval/configs/our_benchmark_config_sub_rxr.yaml --gr00t_port 9000 --output_path /data/sjh/InternNav/output --save_video --num_history 12

Enactive HTTP action contract

internnav/evaluator/HTTPTrajectoryClient.py supports both response forms:

• Legacy Habitat responses with an actions list are executed unchanged.
• Canonical continuous responses use schema_version=2 and carry a full continuous_action[16][4] chunk plus chunk_execute_horizon. The client reconstructs that prefix with canonical_relative_v1 body-frame SE(2) composition, then converts it to native Habitat actions locally.

Each observation advertises client_capabilities (including high_policy_replan_ack_v1) for server-side negotiation and includes executed_actions, the discrete actions successfully executed since the previous HTTP query. STOP remains Habitat action 0; oracle-goal responses keep using the local ShortestPathFollower; control-only replan responses are ACKed and retried without executing an environment action. Protocol/observation errors abort the rollout instead of being interpreted as a normal STOP.

Physical camera directions

The four current RGB views share one agent position. Habitat looks along local -Z; a native TURN_LEFT rotates about +Y. Therefore the left sensor uses +pi/2 yaw relative to the front sensor, right uses -pi/2, and rear uses pi. The request keeps these physical names in rgb_views; it does not exchange left and right later in the transport.

The integration regression constructs the actual Evaluator and Habitat Env, then compares the left/right cameras with the front camera after native 90-degree turns, and the rear camera after a native 180-degree turn. It checks actual sensor rotations, rendered pixels, unchanged position, and the canonical request received from the real client payload builder. Both level cameras and the default two LOOK_DOWN actions are covered. No policy weights or HTTP server are needed.

Run it in the normal Habitat evaluator environment with Enactive importable and a config pointing to real scene/dataset assets:

HABITAT_CAMERA_TEST_CONFIG=/absolute/path/to/habitat.yaml \
HABITAT_CAMERA_TEST_GPU=0 \
PYTHONPATH="$PWD:$PWD/depth_camera_filtering-main" \
python -m pytest tests/integration/test_habitat_camera_directions.py -q

The default split is val_unseen; set HABITAT_CAMERA_TEST_SPLIT when needed. The configured native turn angle must divide 90 degrees. The test skips when HABITAT_CAMERA_TEST_CONFIG is absent; invalid render buffers fail the test.