An action-conditioned latent world model (LeWM) paired with a PRISM action prior, both trained on real Franka FR3 PushT teleoperation data. Supports three planning modes: PRISM-MPPI (PoG fusion, default), warm-start, and vanilla MPPI — selectable at deploy time for A/B comparison.

What's new vs v2

Component	v2	v3
Training dataset	304 eps / 72 k frames	411 eps / 94 k frames
WM val_pred_loss	0.0045	0.0046 (similar)
CV @ H=5	0.180	0.195
pred/id @ H=1	0.465	0.402 (better short-h fidelity)
PRISM action prior	❌ not included	✅ `prior_head_pusht_fr3_v3.pt` (H=3, val MSE 0.348)
PRISM-MPPI mode	❌ vanilla only	✅ 3 modes: pog / warm_start / none
Default `H` (plan-steps)	3	3 (matches both the prior and the WM's deploy sweet spot)

Model summary

	World Model	PRISM Prior Head
Architecture	ViT-tiny encoder + 6-layer AR-Transformer predictor + Embedder action encoder	3-layer MLP, hidden=512, β-NLL loss
Parameters	18.03 M	0.51 M
Input	(224, 224, 3) RGB obs + goal	(z_t, z_g) ∈ ℝ^(2 × 192)
Output	next-latent prediction	(μ, σ) over action sequence (H=3, A_block=5, action_dim=2) — 15 ticks = 1.5 s
Training data	All 411 eps (mixed: 200 target + 211 random terminations)	eps 0-199 only (target-completion subset)
Goal supervision	(n/a — predictor)	HER hindsight, episode last frame (sim convention, Andrychowicz et al. 2017)
Final val_pred_loss / val MSE	0.0046	0.348 (σ ≈ 0.53 ≈ √MSE, well-calibrated)

Why H=3 (not the sim convention H=5)?

The v3 WM's per-step rollout fidelity (pred/id) is best at short horizons (0.40 @ H=1, 0.186 @ H=5, 0.331 @ H=25 — see docs/34). H=3 keeps both the WM rollout and the prior's action sequence within the high-fidelity envelope. Empirically, training the prior at H=3 (vs sim's H=5) gives −8.7 % val MSE (0.348 vs 0.381) and a tighter σ (0.53 vs 0.58) — full ablation in docs/35 §11.

Why train the prior on only the first 200 episodes?

Eps 200-410 were collected with no fixed-target pushing — the operator stopped T at arbitrary positions, making the episode-last-frame z_g a noisy supervision signal. Training on the full 411 eps with HER endframe yields a broken prior (val MSE 1.63, fails HARD GATE). Restricting to the target-clean subset 0-199 recovers a useful prior (val MSE 0.38, well-calibrated σ). See docs/35 §9 for the full ablation.

Plan-worthiness diagnostics (WM only — measured on the train distribution)

Metric @ H=5	Value	Interpretation
CV	0.195 ± 0.010	Borderline below 0.30 plan-worthy threshold; PRISM prior helps
GT_rank	36.6 % ± 2.1	Direction correct ("weak-align" tier)
pred/id @ H=1	0.402 ± 0.016	Good single-step action conditioning
pred/id @ H=5	0.186 ± 0.006	Reasonable 2.5 s rollout fidelity
pred/id @ H=25	0.331 ± 0.006	Long-horizon drift — use H ≤ 5

See docs/34 for the full data-scaling analysis (v1 → v2 → v3 monotone CV climb).

Three planning modes

Mode	Description	When to use
`pog` (default)	PRISM-MPPI: Product-of-Gaussians fusion of prior (μ, σ) into MPPI init	Main deploy mode
`warm_start`	Use prior mean only, keep planner default σ	A/B test (isolates σ-fusion's contribution)
`none`	Vanilla LeWM-MPPI (no prior)	Paper-grade A/B baseline, or no-prior fallback

All three modes share the same LeWM, MPPI loop, K=300, n_iters=30, and action scaler — apples-to-apples comparison.

Quick start

pip install torch torchvision numpy einops transformers huggingface_hub

python

from huggingface_hub import snapshot_download
import numpy as np

# Download the bundle (WM + prior + inference code)
local = snapshot_download("YuhaiW/lewm-pusht-fr3-v3")

import sys; sys.path.insert(0, local)
from pusht_lewm_inference import PushtLewmInference

# ── PRISM-MPPI deploy (recommended default) ─────────────────────────────
planner = PushtLewmInference(
    lewm_ckpt      = f"{local}/lewm_pusht_fr3_v3.ckpt",
    prior_ckpt     = f"{local}/prior_head_pusht_fr3_v3.pt",
    injection_mode = "pog",         # "pog" | "warm_start" | "none"
    device         = "cuda",
)

# In the robot control loop (10 Hz):
while not done:
    obs_uint8  = camera_rgb()             # (224, 224, 3) uint8
    goal_uint8 = goal_rgb()               # (224, 224, 3) uint8
    actions    = planner.plan(obs_uint8, goal_uint8)
                                          # (5, 2) float32 — Δxy meters for next 0.5 s
    for a in actions:
        robot.send_delta_target(a)
        time.sleep(0.1)                   # 10 Hz tick

To A/B test against vanilla MPPI on the same WM:

python

vanilla = PushtLewmInference(
    lewm_ckpt      = f"{local}/lewm_pusht_fr3_v3.ckpt",
    prior_ckpt     = f"{local}/prior_head_pusht_fr3_v3.pt",   # loaded for scaler
    injection_mode = "none",                                  # disable PoG fusion
)

Robot expectations


Robot	Franka FR3 (or compatible) with Cartesian impedance control
Action interpretation	Δ-target XY in meters (per tick)
Control frequency	10 Hz
Camera	Top-down RGB at 224 × 224
Goal image	Single RGB showing the desired final scene
Z, rotation, gripper	NOT controlled (XY-only by design; lock in your controller)
Teleop style assumed	"Decisive" pushes — operator commits and pushes in one smooth motion

What's in the bundle

lewm_pusht_fr3_v3.ckpt                 # 72 MB — world model (pickled JEPA object)
prior_head_pusht_fr3_v3.pt             # 2 MB  — PRISM prior head + StandardScaler + meta
action_scaler.json                     # 0.5 KB — fallback scaler when no prior_ckpt
pusht_lewm_inference.py                # standalone PRISM-MPPI planner (3 modes)
jepa.py, module.py                     # required for LeWM ckpt deserialization
prior_head.py                          # required for prior ckpt deserialization
requirements.txt                       # minimal deps
README.md                              # this file

Architecture overview

   obs (224²)                goal (224²)
        │                          │
        ▼                          ▼
   ViT-tiny ↑                 ViT-tiny ↑       ← shared weights
        │                          │
   z_t ∈ ℝ^192             z_g ∈ ℝ^192

   ─── PRISM Prior Head (optional, "pog" / "warm_start" modes) ──
                  │                  │
                  ▼                  ▼
        concat(z_t, z_g) → MLP → (μ_p, σ_p)  ∈ ℝ^(H × A_block × A_raw)
        │                                       │
        ▼                                       ▼
   ─── PoG fusion with MPPI init μ=0, σ=var_scale ──
                  │
                  ▼
        N(μ_fused, σ_fused) sampled K=300 times
                  │
   ─── MPPI iterations (LeWM AR rollout cost vs z_g) ──
                  │
                  ▼
        optimized action sequence (H × A_block, A_raw)
                  │
                  ▼
        first A_block actions → robot

Caveats and limitations

PRISM prior trained on 200 eps (target subset). The full 411-ep dataset is heterogeneous (eps 200-410 have random T-final-positions). Training the prior on the clean subset gives a usable signal (val MSE 0.38) but still ~3× worse than sim/red-cube counterparts. The next-best improvement would be collecting future datasets with an explicit goal_pixels field (one printed target image per session).
WM cost surface is borderline (CV @ H=5 = 0.195 < 0.30). The PRISM prior is expected to help bridge the gap; PRISM-MPPI's cost-rescoring step tolerates the borderline cost surface in a way that vanilla MPPI cannot.
Trained on top-down RGB only. Other camera angles are OOD.
2-D XY action space. Z, rotation, gripper are not controlled.
10 Hz tick. Faster/slower control loops mismatch the action scaler.

Provenance

WM trained: 2026-06-03 (RTX 5090, ~4 h 53 min on all 411 eps)
Prior trained: 2026-06-03 (RTX 5090, ~30 s on first 200 eps + sim-aligned HER)
Dataset snapshot: Rongxuan-Zhou/pusht_lewm_fr3 sha 1b5dd5db801ef405b43d51dd5b9a3210d8d79ce6
Project: PRISM-JEPA
Companions:
- YuhaiW/lewm-pusht-fr3-v2 — previous release (304 eps, no prior, vanilla MPPI only)
- YuhaiW/prism-jepa-red-cube-arx — same architecture on ARX cube task

Citation

bibtex

@misc{prism-jepa-pusht-fr3-v3,
  title  = {LeWM + PRISM-MPPI for Franka PushT (v3 — 411-ep)},
  author = {Wang, Yuhai and Zhou, Rongxuan and collaborators},
  year   = {2026},
  url    = {https://huggingface.co/YuhaiW/lewm-pusht-fr3-v3}
}

If you cite the PRISM action prior mechanism, also cite Andrychowicz et al. (2017) for hindsight experience replay, on which our prior training is based.

License

Apache 2.0.