24jihoward/openpi-realworld

OpenPI Real-World: Pi0.5 with Track Head + VLAC Reranking Pipeline

Modified OpenPI with real-world Franka Panda support, track-head integration, and a VLAC reranking pipeline for action selection.

Repo Structure

openpi/                          # Modified OpenPI codebase
  src/openpi/
    models/
      pi0.py                    # Pi0 base model (flow matching, joint action+track denoising)
      pi05_vertex_input_asymmetric.py  # Pi0.5 with vertex/track input
    policies/
      realworld_ee_policy.py    # Real-world Franka EE policy adapter
    training/
      config.py                 # All model configs (pi05_realworld_track_joint, etc.)
  checkpoints/
    pi05_realworld_track_joint/run2/99999/   # Actions + 78D track predictions
    pi05_realworld_ee_squarecrop/v2/49999/   # Actions only (square-cropped input)
    pi05_realworld_ee/run1/49999/            # Actions only (padded input)

scripts/vlac_pipeline/           # VLAC reranking pipeline
  openpi_policy.py              # Batched noise candidate generation at diffusion layer
  vlac_rerank_server.py         # Single-GPU: OpenPI + WM + VLAC websocket server
  vlac_rerank_server_multigpu.py # Multi-GPU: process-per-GPU parallel reranking
  vlac_critic.py                # VLAC-8b scoring wrapper
  action_wm.py                  # Action-conditioned SVD world model wrapper
  track_wm.py                   # Track-conditioned SVD+ControlNet world model wrapper
  ctrl_world_wm.py              # Ctrl-World action-conditioned world model wrapper
  track_dualview_wm.py          # Dual-view track-conditioned world model wrapper
  cem_rerank_server.py          # CEM (Cross-Entropy Method) action selection server
  cem_sampler.py                # CEM noise sampling with warm-starting
  state_to_tracks.py            # EE state -> 78D track query points via camera calibration
  VLAC_PIPELINE_PLAN.md         # Pipeline architecture documentation
  CEM_PIPELINE_PLAN.md          # CEM action generation documentation
  SAMPLING_DIVERSITY.md         # Noise sampling at diffusion layer documentation

Models

Config	Checkpoint	Output	Description
pi05_realworld_track_joint	run2/99999	10D actions + 78D tracks	Joint action+track flow matching with track head
pi05_realworld_ee_squarecrop	v2/49999	10D actions	Actions only, square center-crop input
pi05_realworld_ee	run1/49999	10D actions	Actions only, padded input

Action format: 10D absolute EE state [pos(3), rot6d(6), gripper_width(1)]

Track format: 78D = [agentview_mesh(7x2) | wrist_mesh(7x2) | wrist_grid(25x2)]

VLAC Reranking Pipeline

The pipeline generates N stochastic action candidates from pi0.5, renders predicted futures via a world model, and scores them with a VLAC critic to select the best action branch.

Observation (camera + EE state)
    |
    v
Pi0.5 Policy (flow matching, N noise samples batched)
    |
    v
N candidate action chunks (16 steps x 10D)
    |
    v
World Model (SVD video generation per candidate)
    |
    v
N predicted future videos (16 frames)
    |
    v
VLAC Critic (InternVL2, pairwise progress scoring)
    |
    v
argmax -> best action chunk -> execute on robot

Candidate Generation

Candidates are generated by batching N different noise vectors at the flow matching diffusion layer, running a single forward pass through the policy. This is more efficient than N sequential infer() calls (prefix KV cache computed once).

See openpi_policy.py:generate_candidates() and SAMPLING_DIVERSITY.md.

Server Architecture

• Single GPU (vlac_rerank_server.py): All models on one GPU, candidates scored serially
• Multi GPU (vlac_rerank_server_multigpu.py): Policy on GPU 0, WM+VLAC workers as separate processes on other GPUs. True parallelism (no GIL).
• CEM (cem_rerank_server.py): Cross-Entropy Method iteratively refines the noise distribution

Documentation

• openpi/PI05_TRACK_PROGRESS.md — Track head integration progress
• openpi/REALWORLD_TRACK_TRAINING.md — Real-world training setup and parameters
• scripts/vlac_pipeline/VLAC_PIPELINE_PLAN.md — Full pipeline architecture, image resolution map, action format conversions
• scripts/vlac_pipeline/CEM_PIPELINE_PLAN.md — CEM action generation design
• scripts/vlac_pipeline/SAMPLING_DIVERSITY.md — Noise sampling strategies

Usage

# Load pi05_realworld_track_joint
from openpi.training import config as _config
from openpi.policies import policy_config as _policy_config

cfg = _config.get_config("pi05_realworld_track_joint")
policy = _policy_config.create_trained_policy(cfg, "openpi/checkpoints/pi05_realworld_track_joint/run2/99999")
result = policy.infer(observation, return_tracks=True)
# result["actions"]: (16, 10) — 10D absolute EE actions
# result["track_predictions"]: (16, 78) — 78D track predictions