Add Env EB-Alfred

examples/evaluate/eb_alfred/config.yaml

@@ -0,0 +1,91 @@
+fileroot: ${oc.env:HOME}/projects/vagen
+
+envs:
+  - name: RemoteEnv
+    n_envs: 50
+    data_source: eb_alfred
+    tag_id: eb_alfred_val_common
+    seed: [0, 50, 1]
+    max_turns: 6
+    config:
+      base_urls:
+        - "http://localhost:8000"
+      timeout: 600
+      eval_set: common_sense
+      obs_image_size: 500
+      max_turns: 6
+      max_actions_per_step: 20
+      max_env_steps: 30
+      action_sep: ","
+      prompt_format: free_think
+      use_example_in_sys_prompt: true
+      format_reward: 0.1
+      success_reward: 1.0
+    chat_config:
+      temperature: 0
+      max_tokens: 2048
+      top_p: 1.0
+
+  - name: RemoteEnv
+    n_envs: 50
+    data_source: eb_alfred
+    tag_id: eb_alfred_val_spatial
+    seed: [0, 50, 1]
+    max_turns: 6
+    config:
+      base_urls:
+        - "http://localhost:8000"
+      timeout: 600
+      eval_set: spatial
+      obs_image_size: 500
+      max_turns: 6
+      max_actions_per_step: 20
+      max_env_steps: 30
+      action_sep: ","
+      prompt_format: free_think
+      use_example_in_sys_prompt: true
+      format_reward: 0.1
+      success_reward: 1.0
+    chat_config:
+      temperature: 0
+      max_tokens: 2048
+      top_p: 1.0
+
+experiment:
+  dump_dir: ${fileroot}/rollouts/eval_eb_alfred
+  default_max_turns: 6
+
+run:
+  backend: "openai"
+  base_seed: 0
+  max_concurrent_jobs: 4
+  resume: skip_completed
+  live_summary: true
+
+backends:
+  openai:
+    api_key: ""               # or env OPENAI_API_KEY
+    base_url: null
+    model: "gpt-4o-mini"
+    max_concurrency: 2
+    max_retries: 6
+    min_backoff: 0.5
+    max_backoff: 8.0
+
+  sglang:
+    base_url: "http://127.0.0.1:30000/v1"
+    api_key: "EMPTY"
+    model: "Qwen/Qwen2.5-VL-7B-Instruct"
+    max_concurrency: 2
+    max_retries: 6
+    min_backoff: 0.5
+    max_backoff: 8.0
+
+  claude:
+    api_key: ""
+    base_url: null
+    model: "claude-3-5-sonnet-latest"
+    max_concurrency: 2
+    max_retries: 6
+    min_backoff: 0.5
+    max_backoff: 8.0

examples/evaluate/eb_alfred/run_eval.sh

@@ -0,0 +1,14 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Before running, start the eb_alfred server in another terminal:
+#   python -m vagen.envs.eb_alfred.serve
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+CONFIG="${1:-$SCRIPT_DIR/config.yaml}"
+shift 2>/dev/null || true
+
+LOG_FILE="run.log"
+
+python -m vagen.evaluate.run_eval --config "$CONFIG" "$@" \
+  2>&1 | tee "${LOG_FILE}"

examples/train/eb_alfred/train_eb_alfred_vision.yaml

@@ -0,0 +1,66 @@
+envs:
+  - name: RemoteEnv
+    n_envs: 50
+    data_source: eb_alfred
+    tag_id: eb_alfred_train_base
+    seed: [0, 200, 1]
+    max_turns: 6
+    response_length_per_turn: 512
+    config:
+      base_urls:
+        - "http://localhost:8000"
+      timeout: 600
+      eval_set: base
+      obs_image_size: 500
+      max_turns: 6
+      max_actions_per_step: 20
+      max_env_steps: 30
+      action_sep: ","
+      prompt_format: free_think
+      use_example_in_sys_prompt: true
+      format_reward: 0.1
+      success_reward: 1.0
+
+  - name: RemoteEnv
+    n_envs: 50
+    data_source: eb_alfred
+    tag_id: eb_alfred_train_complex
+    seed: [0, 200, 1]
+    max_turns: 6
+    response_length_per_turn: 512
+    config:
+      base_urls:
+        - "http://localhost:8000"
+      timeout: 600
+      eval_set: complex_instruction
+      obs_image_size: 500
+      max_turns: 6
+      max_actions_per_step: 20
+      max_env_steps: 30
+      action_sep: ","
+      prompt_format: free_think
+      use_example_in_sys_prompt: true
+      format_reward: 0.1
+      success_reward: 1.0
+
+  - name: RemoteEnv
+    n_envs: 50
+    data_source: eb_alfred
+    tag_id: eb_alfred_train_visual
+    seed: [0, 200, 1]
+    max_turns: 6
+    response_length_per_turn: 512
+    config:
+      base_urls:
+        - "http://localhost:8000"
+      timeout: 600
+      eval_set: visual_appearance
+      obs_image_size: 500
+      max_turns: 6
+      max_actions_per_step: 20
+      max_env_steps: 30
+      action_sep: ","
+      prompt_format: free_think
+      use_example_in_sys_prompt: true
+      format_reward: 0.1
+      success_reward: 1.0

examples/train/eb_alfred/train_ppo_no_concat_qwen25vl3b.sh

@@ -0,0 +1,86 @@
+#!/bin/bash
+
+set -x
+
+PROJECT_NAME="vagen_experiments"
+EXPERIMENT_NAME="ppo_eb_alfred_no_concat"
+
+BASEDIR=$(pwd)
+SCRIPTDIR=$(dirname "$0")
+EXPERIMENT_DIR=${BASEDIR}/exps/${PROJECT_NAME}/${EXPERIMENT_NAME}
+SAVE_CHECKPOINT_DIR=${EXPERIMENT_DIR}/verl_checkpoints
+DATASET_TRAIN=${SCRIPTDIR}/train_eb_alfred_vision.yaml
+DATASET_VAL=${SCRIPTDIR}/val_eb_alfred_vision.yaml
+agent_loop_config_path=${BASEDIR}/vagen/configs/agent_no_concat.yaml
+REF_MODEL_PATH=Qwen/Qwen2.5-VL-3B-Instruct
+mkdir -p ${EXPERIMENT_DIR}
+
+export HF_HOME=/workspace/.hf_home
+export PATH=/venv/vagen/bin:$PATH
+
+PYTHONUNBUFFERED=1 python3 -m vagen.main_ppo \
+    --config-path=${BASEDIR}/vagen/configs \
+    --config-name='vagen_multiturn' \
+    data.train_files=${DATASET_TRAIN} \
+    data.val_files=${DATASET_VAL} \
+    data.train_batch_size=128 \
+    data.max_prompt_length=2048 \
+    data.max_response_length=512 \
+    +data.max_trajectory_length=7000 \
+    algorithm.adv_estimator=no_concat_gae_first \
+    algorithm.kl_ctrl.kl_coef=0.0 \
+    actor_rollout_ref.model.path=${REF_MODEL_PATH} \
+    actor_rollout_ref.model.use_remove_padding=True \
+    actor_rollout_ref.model.use_fused_kernels=True \
+    actor_rollout_ref.model.enable_gradient_checkpointing=True \
+    actor_rollout_ref.actor.optim.lr=1e-6 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=32 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
+    actor_rollout_ref.actor.use_kl_loss=False \
+    actor_rollout_ref.actor.kl_loss_coef=0.0 \
+    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+    actor_rollout_ref.actor.entropy_coeff=0.0 \
+    actor_rollout_ref.actor.checkpoint.save_contents=['model','hf_model','optimizer','extra'] \
+    actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
+    actor_rollout_ref.actor.fsdp_config.param_offload=True \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+    actor_rollout_ref.actor.freeze_vision_tower=True \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
+    actor_rollout_ref.rollout.name=sglang \
+    actor_rollout_ref.rollout.mode=async \
+    actor_rollout_ref.rollout.n=1 \
+    actor_rollout_ref.rollout.max_num_batched_tokens=10000 \
+    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
+    actor_rollout_ref.rollout.enforce_eager=True \
+    actor_rollout_ref.rollout.free_cache_engine=True \
+    actor_rollout_ref.rollout.enable_chunked_prefill=True \
+    actor_rollout_ref.rollout.multi_turn.enable=True \
+    actor_rollout_ref.rollout.agent.agent_loop_config_path=$agent_loop_config_path \
+    actor_rollout_ref.rollout.disable_log_stats=False \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
+    actor_rollout_ref.ref.fsdp_config.param_offload=True \
+    critic.enable=True \
+    critic.optim.lr=1e-5 \
+    critic.model.use_remove_padding=True \
+    critic.model.path=${REF_MODEL_PATH} \
+    critic.model.enable_gradient_checkpointing=True \
+    critic.ppo_micro_batch_size_per_gpu=1 \
+    critic.model.fsdp_config.param_offload=True \
+    critic.model.fsdp_config.optimizer_offload=True \
+    trainer.critic_warmup=0 \
+    trainer.logger=['console','wandb'] \
+    trainer.val_before_train=True \
+    trainer.n_gpus_per_node=4 \
+    trainer.nnodes=1 \
+    trainer.save_freq=100 \
+    trainer.test_freq=20 \
+    trainer.project_name=${PROJECT_NAME} \
+    trainer.experiment_name=${EXPERIMENT_NAME} \
+    trainer.default_local_dir=${SAVE_CHECKPOINT_DIR} \
+    trainer.validation_data_dir=${EXPERIMENT_DIR}/validation \
+    trainer.rollout_data_dir=${EXPERIMENT_DIR}/rollout_data \
+    trainer.log_val_generations=32 \
+    +trainer.concat_multi_turn=False \
+    trainer.total_training_steps=400 2>&1 | \
+    tee ${EXPERIMENT_DIR}/${PROJECT_NAME}_${EXPERIMENT_NAME}.log >(tee ${BASEDIR}/${PROJECT_NAME}_${EXPERIMENT_NAME}.log >/dev/null)

examples/train/eb_alfred/val_eb_alfred_vision.yaml

@@ -0,0 +1,44 @@
+envs:
+  - name: RemoteEnv
+    n_envs: 50
+    data_source: eb_alfred
+    tag_id: eb_alfred_val_common
+    seed: [0, 50, 1]
+    max_turns: 6
+    response_length_per_turn: 512
+    config:
+      base_urls:
+        - "http://localhost:8000"
+      timeout: 600
+      eval_set: common_sense
+      obs_image_size: 500
+      max_turns: 6
+      max_actions_per_step: 20
+      max_env_steps: 30
+      action_sep: ","
+      prompt_format: free_think
+      use_example_in_sys_prompt: true
+      format_reward: 0.1
+      success_reward: 1.0
+
+  - name: RemoteEnv
+    n_envs: 50
+    data_source: eb_alfred
+    tag_id: eb_alfred_val_spatial
+    seed: [0, 50, 1]
+    max_turns: 6
+    response_length_per_turn: 512
+    config:
+      base_urls:
+        - "http://localhost:8000"
+      timeout: 600
+      eval_set: spatial
+      obs_image_size: 500
+      max_turns: 6
+      max_actions_per_step: 20
+      max_env_steps: 30
+      action_sep: ","
+      prompt_format: free_think
+      use_example_in_sys_prompt: true
+      format_reward: 0.1
+      success_reward: 1.0

vagen/envs/eb_alfred/README.md

@@ -0,0 +1,93 @@
+# EB-ALFRED Environment
+
+AI2-THOR based household robot task environment from [EmbodiedBench](https://github.com/EmbodiedBench/EmbodiedBench). The agent receives egocentric RGB images and executes multi-step tasks (cleaning, heating, slicing, storing objects).
+
+## Running the Service
+
+The environment runs on a **separate GPU machine** with a physical or virtual display. AI2-THOR requires X11 rendering (CloudRendering is not supported on ai2thor 2.1.0).
+
+**One-time setup — create the conda environment:**
+
+```bash
+conda create -n embodiedbench python=3.9 -y
+conda activate embodiedbench
+
+git clone https://github.com/EmbodiedBench/EmbodiedBench.git /root/EmbodiedBench
+touch /root/EmbodiedBench/embodiedbench/__init__.py
+pip install -e /root/EmbodiedBench
+
+pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
+pip install "ai2thor==2.1.0" "gym==0.23.0" "numpy<2.0" \
+    scipy Pillow networkx revtok vocab h5py tqdm natsort pyquaternion
+pip install "flask==1.1.4" "werkzeug==1.0.1" \
+    "markupsafe<2.1" "jinja2<3.0" "itsdangerous<2.0"
+pip install "opencv-python-headless<4.9"
+```
+
+**Download dataset** (`eval_set` selects the split: `base` — standard tasks, `long` — longer horizon):
+
+```bash
+git clone https://huggingface.co/datasets/EmbodiedBench/EB-ALFRED
+mv EB-ALFRED /root/EmbodiedBench/embodiedbench/envs/eb_alfred/data/json_2.1.0
+```
+
+**Start the server** (GPUs and Xorg are auto-detected and started):
+
+```bash
+conda activate embodiedbench
+python -m vagen.envs.eb_alfred.serve
+```
+
+Key parameters:
+- `devices`: GPU indices (default: auto-detect via `CUDA_VISIBLE_DEVICES` or `nvidia-smi`)
+- `capacity`: max concurrent Unity environments (default: 16)
+- `startup_concurrency`: max Unity processes starting simultaneously, prevents CPU spikes (default: 8)
+- `session_timeout`: idle session cleanup in seconds (default: 3600)
+
+```bash
+# Example: 2 GPUs, higher capacity
+python -m vagen.envs.eb_alfred.serve --devices='[0,1]' --capacity=90 --startup_concurrency=6 --port=8000
+```
+
+**SSH tunnel** (if training machine is remote):
+
+```bash
+# Run on the env server — forwards port 8000 to training machine
+ssh -p <PORT> -R 8000:localhost:8000 \
+    -o ServerAliveInterval=30 -o ServerAliveCountMax=5 \
+    -N -f user@training-machine-ip
+
+# On training machine, allow many tunnels — add to /etc/ssh/sshd_config:
+# MaxSessions 200
+# then: service ssh reload
+```
+
+## Evaluation
+
+```bash
+conda activate vagen
+
+# Terminal 1 (env server): start service
+python -m vagen.envs.eb_alfred.serve --devices='[0,1]' --capacity=90
+
+# Terminal 2 (training machine): run eval
+python -m vagen.evaluate.run_eval --config examples/evaluate/eb_alfred/config.yaml
+```
+
+Config: `examples/evaluate/eb_alfred/config.yaml`
+
+## Training
+
+```bash
+conda activate vagen
+
+# Terminal 1 (env server): start service
+python -m vagen.envs.eb_alfred.serve --devices='[0,1]' --capacity=90
+
+# Terminal 2 (training machine): run training
+cd VAGEN
+bash examples/train/eb_alfred/train_grpo_qwen25vl3b.sh
+```
+
+Configs: `examples/train/eb_alfred/`
+