[Refactor] Habitat Refactor & Support Distributed VLN-PE Evaluation

[kew6688]

PR Summary

• Refactored Habitat integration into the InternNav framework for improved readability and reusability.
  • Maintained a unified evaluation script for all InternNav benchmarks with the same config format and starter workflow.
  • Created a reusable Distributed Evaluator that encapsulates distribution logic for multiple benchmarks.
  • Added a Habitat Env wrapper that handles reset/step, episode assignment, and episode reloading inside the worker, significantly streamlining the evaluation pipeline.
• Refactored VLNPE evaluation flow:
  • Added an option to bypass the agent server for models that cannot process batched observations, splitting evaluation per distributed rank.
  • Now VlnMultiEvaluator inherited from DistributedEvaluator, the same to Habitat. Ensured each Isaac Sim simulator and model instance is bound to a dedicated GPU, aligning with the Habitat evaluator’s design and enabling deployment on both Aliyun DLC and Slurm.
  • refactor new episode loader and generate all episodes under internutopia env. Keep the evaluator logic clean and align with Habitat evaluation workflow.

Distributed Eval Time

Using 16 nodes, each with 1× RTX 4090 GPU, 8 CPUs, and 60 GB RAM.

Model Variant	GPUs Used	Previous Runtime (Single 4090)	Distributed Runtime (16×4090)	Speedup
InternVLA-N1 Flash	1 GPU → 16 GPUs	~13.5 hours	~1 hour	13× faster
InternVLA-N1 PE`	1 GPU → 16 GPUs	~21 hours	~1.6 hour	13× faster

Performance after Refactor

• Re-ran InternNav benchmarks (Internutopia VLNPE + Habitat VLNCE) three times post-refactor, matching previous performance.

Model	Dataset/Benchmark	NE	OS	SR	SPL
InternVLA-N1	Habitat R2R	4.88	62.2	57.0	52.4
InternVLA-N1	Flash	4.17	67.2	59.8	53.9
InternVLA-N1	PE	4.87	55.7	50.0	42.9
RDP	Flash	7.11	41.7	24.3	17.4
RDP	PE	6.73	38.0	26.3	18.6

[yuqiang-yang]

这个是干什么的，habitat vln

[kew6688]

这个flag没有用到，我去掉

[yuqiang-yang]

我看server里面写了默认的host和port，这里也写一下会不会好一些，方便跑

[kew6688]

好的，已在AgentConfig里面加了默认的host和port