#Policies

Three policy groups run in closed loop (paper §4.1, App. D): VLAs, and VLMs driving VoxPoser [1] or PIVOT [2]. Commands run in the container (Getting started) with the repository at /workspace; blocking_tasks.txt holds the 125 benchmark classes (Benchmark).

#VLA policies

Released RoboCasa checkpoints of four VLAs, weights fixed. Each is a profile in policy/VLA/src/configs/models.yaml:

• Input: instruction c1 (Experiments), left, right and wrist RGB at 256 × 256, and proprioception; no top-down image, obstacle position or tier label.
• Output: 12-D actions through the HYBRID_MOBILE_BASE composite controller with --base-controller legacy, at 20 Hz.

Serve each model from its own environment (policy/VLA/README.md §2.2; GR00T N1.5 with the groot_n1_5_atomic_seen profile). One seed of one model:

python3 policy/VLA/src/run_vla_eval.py \
  --model-profile xr1 --policy-url http://127.0.0.1:8100 \
  --tasks $(cat blocking_tasks.txt) --layout-ids 0,2,5,7,8 --style MODERN_1 \
  --seed 0 --policy-seed 0 --horizon 500 \
  --out-dir /workspace/outputs/xr1_seed0 --ledger-dir /workspace/outputs/xr1_seed0/ledger

--num-workers N --worker-id i splits a seed over workers. Without --tasks the runner also takes navigation classes outside OSMA-Bench.

#VLM-based policies

A runner reaches the model's OpenAI-compatible endpoint through -m <served model name> -p <port>.

• VoxPoser [1] (policy/Voxposer, code generation): the VLM writes code that builds value maps and sets obstacle-specific clearance and velocity; an A* planner turns the maps into a path.
• PIVOT [2] (policy/PIVOT, waypoint selection): the VLM picks the next waypoint among candidate points marked on the scene image, over several rounds.

Both drive the base through a shared holonomic waypoint follower (robosuite's 11-D BASIC action). NAV_MODE=holonomic selects it over navigation_controller.mode in policy/Voxposer/src/configs/robocasa_config.yaml, and run.json records it as nav_mode.

One seed of one model. Worker i runs a slice of blocking_tasks.txt; the folder names are what scripts/merge_seed_shards.py looks for:

OUT=/workspace/outputs; M=<model tag>; S=0; i=0; TASKS="<slice i of blocking_tasks.txt>"

# VoxPoser
cd /workspace/policy/Voxposer && MUJOCO_GL=egl NAV_MODE=holonomic \
  ROBOCASA_SEED=$S ROBOCASA_LEDGER_DIR=$OUT/${M}_seed${S}_vox_s${i}_ledger \
  python3 src/run_LMP.py $TASKS -m <served model name> -p <port> \
    --layout-ids 0,2,5,7,8 --style-ids 3 --seed $S -w $i -o $OUT/${M}_seed${S}_vox_s${i}

# PIVOT
cd /workspace/policy/PIVOT && MUJOCO_GL=egl NAV_MODE=holonomic \
  ROBOCASA_SEED=$S ROBOCASA_LEDGER_DIR=$OUT/${M}_seed${S}_piv_s${i}_ledger \
  python3 run_pivot.py $TASKS -m <served model name> -p <port> \
    --layout-ids 0,2,5,7,8 --style-ids 3 --seed $S --subgoals pivot -o $OUT/${M}_seed${S}_piv_s${i}

# after all workers finish: one folder per model, seed and policy
python3 /workspace/scripts/merge_seed_shards.py --root $OUT --prefix $M --seed $S \
  --suffix vox --dest $OUT/${M}_seed${S}_voxposer
python3 /workspace/scripts/merge_seed_shards.py --root $OUT --prefix $M --seed $S \
  --suffix piv --dest $OUT/${M}_seed${S}_pivot

• ROBOCASA_SEED sets the environment seed (42 when unset) and --seed the policy's sampling seed; the paper sets both to the same value.
• PIVOT needs --layout-ids and --subgoals pivot; its defaults are layout 0 and solver subgoals.

#All policies

• A task ends at task success or when the step budget runs out: --horizon control steps for VLAs, max_steps_per_waypoint (robocasa_config.yaml) per waypoint for the planners.
• Re-running a command with the same output folder (and worker id, where there is one) skips the tasks it has already finished.
• Score the ledgers with Metrics.

#References

1. W. Huang et al. VoxPoser: Composable 3D value maps for robotic manipulation with language models. CoRL, 2023.
2. S. Nasiriany et al. PIVOT: Iterative visual prompting elicits actionable knowledge for VLMs. ICML, 2024.
3. Xiaomi Robotics Team et al. Xiaomi-Robotics-1: Scaling vision-language-action models with over 100k hours of real-world trajectories. arXiv:2607.15330, 2026.
4. Tensor. OpenTau: Tensor's VLA training infrastructure for real-world robotics in PyTorch. https://github.com/TensorAuto/OpenTau, 2025.
5. NVIDIA et al. GR00T N1: An open foundation model for generalist humanoid robots. arXiv:2503.14734, 2025.
6. NVIDIA. GR00T N1.5: An improved open foundation model for generalist humanoid robots. https://research.nvidia.com/labs/gear/gr00t-n1_5/, 2025.
7. Physical Intelligence et al. π0.5: A vision-language-action model with open-world generalization. CoRL, 2025.
8. S. Bai et al. Qwen3-VL technical report. arXiv:2511.21631, 2025.
9. Gemma Team. Gemma 4 technical report. arXiv:2607.02770, 2026.
10. NVIDIA et al. Cosmos-Reason1: From physical common sense to embodied reasoning. arXiv:2503.15558, 2025.
11. H. Tan et al. RoboBrain 2.5: Depth in sight, time in mind. arXiv:2601.14352, 2026.