morealcholplz/vlsa-dual-safety-eval

VLSA Dual-Arm Safety Evaluation (AEGIS, dual-arm extension)

Quantitative evaluation of a plug-and-play CBF safety layer (VLSA/AEGIS) extended to a dual-arm setting, where two Franka Panda arms face each other across a table and each treats the opposing arm's end-effector as a dynamic ellipsoidal obstacle.

This is a dual-arm extension of VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer (AEGIS, arXiv:2512.11891). The original work is single-arm with static obstacles on a pretrained π0.5 policy; here the policies are self-trained SmolVLA (fine-tuned from lerobot/smolvla_base) and the safety threat is the other moving arm.

Dataset viewer: the results config (per-episode evaluation records, one CSV per split) renders in the viewer. The videos/ folder holds paired baseline-vs-safe rollout MP4s. For an interactive trajectory viewer (LeRobot format), see the companion dataset of recorded rollouts.

Task & setup

• Env: LIBERO_Dual_Tabletop_Manipulation (SafeLIBERO / robosuite), 2 opposing Panda arms.
• Task: place milk into the target region inside a central trash can. Success is scored on robot0's object (milk_1).
• Policy: SmolVLA (joint-space action: 7 joint deltas + gripper), 3 cameras + state(6).
• Safety layer (CBF): each safety-enabled arm models its own and the opposing arm's EE as ellipsoids (semi-axes 0.06/0.12/0.11 m + margin), forms a control barrier function on the center-line directional distance, and minimally corrects the EE displacement via a half-space projection, then maps the correction back to joint space with a damped Jacobian pseudo-inverse (action calibrated scale=0.166). No QP solver needed (single linear constraint, closed form). The policy's joint intent is preserved; the layer activates only on imminent violation.
• Collision metric: any mujoco contact between a robot0 geom and a robot1 geom.

Results — VLSA matrix (robot0_v4 / robot1_v4, 50 seeds × 1 trial, max-steps 400)

Condition	safety arms	TSR	CAR (overall)	CAR (policy-phase)	avg interventions
VLSA-0 (none)	–	4%	4%	16%	0
VLSA-1 (arm0)	[0]	0%	42%	54%	28.9
VLSA-2 (both)	[0,1]	0%	64%	76%	50.1

• CAR (collision avoidance) increases monotonically with the number of safety arms: 4% → 42% → 64% (VLSA-2 ≈ 16× the baseline). Policy-phase CAR (collisions after both arms are under policy control, which the CBF actually governs): 16% → 54% → 76%. This confirms the paper's core hypothesis in the dual-arm regime.
• VLSA-1 is partial/asymmetric: a single yielding arm cannot avoid an actively-approaching one in a head-on shared-goal task; both arms need the layer (VLSA-2).
• TSR trade-off (→0%): both arms must place into the same central trash can, so strong mutual avoidance prevents the close central approach the task requires — the paper's "safety-induced distribution shift", amplified here by the shared-goal geometry. The single-arm ceiling (ceiling_robot0_alone) is itself only ~10%, i.e. TSR is limited by the SmolVLA policy, not by collisions.

Note: SmolVLA action generation is stochastic (flow-matching). These numbers use 1 trial per seed. A per-episode-seeded reproducible re-run is available alongside.

Contents

README.md                      this card
SAFETY_EXPERIMENT_LOG.md       full methodology + timeline log
code/
  cbf_safety.py                dual-arm CBF safety layer
  run_safety_eval.py           multi-seed eval harness (TSR/CAR aggregation, video, LeRobot record)
results/
  baseline_v4.{csv,json}       VLSA-0
  vlsa1_v4.{csv,json}          VLSA-1 (arm0)
  vlsa2_v4.{csv,json}          VLSA-2 (both)
  ceil_r0_v4.{csv,json}        robot0 single-arm TSR ceiling
videos/
  VLSA0_v4_seed{1,3,8,26}.mp4  baseline rollouts (collisions)
  VLSA2_v4_seed{1,3,8,26}.mp4  with both-arm CBF (collision avoidance)

Each results/*.csv is per-episode (seed, success, collided, collision_step, n_intervene, min_h, ...); the matching *.json adds an aggregate summary.

Reproduce

Code is also on GitHub (minje227-coder/RobotLearning_9, branch hannuri). Example:

python run_safety_eval.py \
  --robot0-policy-path <robot0_v4>/checkpoints/last/pretrained_model \
  --robot1-policy-path <robot1_v4>/checkpoints/last/pretrained_model \
  --seed-start 1 --num-episodes 50 --max-steps 400 \
  --safety-arms 0 1 \
  --out-csv results/vlsa2.csv --out-json results/vlsa2.json

Citation

Original method: VLSA / AEGIS — Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer, arXiv:2512.11891.