INSAIT-Institute/LADDER-Bridge-Policy

LADDER Bridge policy

📄 Paper · 💻 Code · 🌐 Website · 🤗 LADDER collection

The WidowX low-level policy of LADDER, a benchmark for long-horizon robot manipulation. It is a pi0-style policy with a Qwen3-VL-2B backbone and a flow-matching action expert, trained on BridgeData v2.

On the two WidowX suites, SimplerEnv-Long and RobotArena-∞-Long, this is the policy that moves the robot:

• The method Low-level policy (low_level_policy) is this policy on its own, given the whole instruction.
• The orchestrator methods (for example Proprioceptive verifier and Privileged verifier) run this same policy and give it one subgoal at a time.

On RoboLab-Long (Franka) the low-level policy is pi0.5 instead.

Files

• checkpoints/model-checkpoint-ema-030000.pt: the weights (EMA, step 30,000).
• hf_export/ladder-bridge-policy/src/: the model code and its configuration.
• SHA256SUMS: every file's checksum (sha256sum -c SHA256SUMS in the downloaded directory).

Usage

With the LADDER code, scripts/eval.sh downloads the policy into the Hugging Face cache on its first run:

scripts/eval.sh --suite simplerenv_long --method low_level_policy

To download it on its own:

hf download INSAIT-Institute/LADDER-Bridge-Policy

Model

architecture	pi0-style: a vision-language backbone and a flow-matching action expert (28 blocks, 10 denoising steps at inference)
backbone	Qwen/Qwen3-VL-2B-Instruct
training data	BridgeData v2 (WidowX 250), from Open X-Embodiment
training	30,000 steps, batch 1,024, AdamW (lr 5e-5, 250 warmup steps, no weight decay, gradient clip 1.0)
checkpoint	EMA of the weights (decay 0.99, from step 20,000) at step 30,000
input	the main camera image, the language instruction and the 8-D robot state
output	a chunk of 8-D actions

Results on LADDER

Task success rate (%) over all tasks of a suite, from the paper. "Low-level policy" is this policy on its own; the verifiers run it one subgoal at a time.

method	SimplerEnv-Long	RobotArena-∞-Long
Low-level policy	8.7	10.9
Proprioceptive verifier	25.2	24.2
Privileged verifier (reference)	29.5	25.5

Limitations

The policy was trained on real WidowX data only and evaluated in the SimplerEnv and RobotArena-∞ digital twins. It was not tuned for other robots or scenes, and it is not a safety-rated controller.

License

MIT. The backbone, Qwen3-VL-2B-Instruct, is released under Apache-2.0, and BridgeData v2 under CC BY 4.0.

Citation

@article{ladder2026,
  title   = {{LADDER}: A Long-Horizon Robot Manipulation Benchmark},
  author  = {Petkov, Petko and Albanese, Giuliano and Nikolov, Nikolay and Van Gool, Luc and Tremblay, Jonathan and Paudel, Danda Pani},
  journal = {arXiv preprint arXiv:2610.18053},
  year    = {2026},
  url     = {https://arxiv.org/abs/2610.18053}
}