tayalmanan/SafeVLA-HJ-Checkpoints

SafeVLA HJ-Reachability Checkpoints

Feasibility-Gated PPO checkpoints with Hamilton-Jacobi reachability cost critic, trained on Safety-CHORES benchmark.

Checkpoints

Checkpoint	Task	Cost Type	Steps	Eval SR	Eval CC
hj_binary_pickup_204K.pt	PickupType	Binary (+25/-1)	204K	0.906	0.25
hj_vlm_rawadv_pickup_462K.pt	PickupType	VLM (rubrics)	462K	0.818	0.52
hj_vlm_fetch_310K.pt	FetchType	VLM (rubrics)	310K	0.515	4.79

Comparison with Baselines

Method	Pickup SR	Pickup CC	Fetch SR	Fetch CC
Lagrangian (ISA, paper)	0.875	0.25	0.637	8.08
HJ-Binary	0.906	0.25	-	-
HJ-VLM (ours)	0.818	0.52	0.515	4.79

Architecture

• Base model: SPOC-DINOv2 (56M trainable params)
• Cost critic: Separate transformer with HJ max-based Bellman backup
• VLM cost scorer: Qwen3-VL-2B-Instruct (rubrics-based, 5 safety dimensions)
• Feasibility gate: Hard binary constraint via cost value function

Key Findings

• Binary +25/-1 costs cause extreme V_c predictions leading to aggressive gate closure
• VLM-calibrated costs (safe=-1, unsafe=[0,25]) provide smoother cost landscape
• FetchType benefits most from HJ gating (CC reduced 41% vs Lagrangian)
• Cost advantage normalization must be removed for correct safety recovery gradients