VLM4VLA — InternVL3.5-1B fine-tuned on Bridge
Vision-Language-Action (VLA) checkpoints from the VLM4VLA framework
(RoboInternVL35 policy + FCDecoder continuous-action head), fine-tuned on the BridgeData V2 (bridge_orig)
manipulation dataset and evaluated in SimplerEnv on the 4 WidowX bridge tasks.
This repository archives two training runs plus the VLM backbone they were fine-tuned from.
Backbone
All runs fine-tune from InternVL3_5-1B-mix_base_embodied_5_0221_proximity_2560k-s10000 — an InternVL3.5-1B
(~1.06B params) variant pre-trained on an embodied mixture (proximity, 2560k samples, snapshot s10000).
The full backbone is included under vlm_backbone/ for reproducibility.
Contents
| Path | Description |
|---|---|
vlm_backbone/ |
The InternVL3.5-1B backbone the runs were fine-tuned from (~2 GB) |
bridge_16k/ |
16k-step run — 8 fp32 checkpoints (step-step=00002000.pt … 00016000.pt), config.json, eval_results/ |
bridge_32k/ |
32k-step run (stopped ~step 19.5k) — 9 fp32 checkpoints (… 00018000.pt), config.json, eval_results/ |
RESULTS.md |
Full per-checkpoint × per-exec-step SimplerEnv-Bridge success-rate tables |
Each checkpoint is a single fp32 .pt state dict (~4 GB) holding the full model (backbone + action head),
converted from DeepSpeed Stage-2 shards via zero_to_fp32.
Training setup (both runs)
- Data:
bridge_orig(BridgeData V2) via the OpenVLA RLDS loader, image aug on, 224×224, window=1,fwd_pred_next_n=4 - Optimizer: Adam, lr 5e-5, 2000 warmup steps, weight decay 0, grad clip 1.0, bf16, DeepSpeed Stage-2
- Global batch size 256 (8×H100, per-GPU 32),
train_vision=true - Action head:
FCDecoder, continuous 7-DoF, action normalized to [-0.65, 0.65] - The two runs differ only in
max_steps(16000 vs 32000), which changes the LR-decay horizon.
Headline results (SimplerEnv-Bridge, AVG over 4 tasks)
| Run | Best config | Success |
|---|---|---|
| 16k | step 16000, exec 4 | 50.0% |
| 32k | step 14000, exec 4 | 51.0% |
exec = number of action steps executed per VLM inference call. See RESULTS.md for the full grid.
Eval is deterministic (fixed episode seeds + bf16), so repeated runs reproduce identical numbers (std ≈ 0).
Usage
These are VLA policy checkpoints intended for use with the VLM4VLA / RoboVLMs codebase. Load a .pt state dict
into the RoboInternVL35 model built from the matching config.json, with vlm_backbone/ as the VLM. See the
framework repo for the inference / SimplerEnv evaluation entry points.
License
The InternVL3.5 backbone follows its upstream license; the VLA wrapper/head follow the VLM4VLA project license.
Marked other — consult the respective upstream repositories before redistribution or commercial use.