Names that appear in configs, launchers, run names and analysis code, and what
they mean in the paper's terms.
| Name in code |
Paper name |
Meaning |
pnp, pnp_cl, PnPRetrieval |
FAR retriever |
the learned, action-conditioned retrieval memory (src/far/memories/pnp_cl.py); "cl" = trained with a contrastive pre-training stage |
meta, pnp_meta, meta_only |
FAR -- Meta |
the retriever scoring past frames with the closed-form metadata cross-scorer only (relative pose, time offsets) |
visual, pnp_visual |
FAR -- Visual |
the retriever scoring with the visual (QueryViT) channel only |
vmz, visual+meta_z |
FAR -- Multi-Cue (LoopNav, AI2-THOR) |
visual + metadata channels, each standardised per query and fused by a learned gate |
amz, cue_metaz, audio+meta_z |
FAR -- Multi-Cue (SoundSpaces) |
the same fusion with the binaural-audio channel in place of the visual one |
ablz, pnp_ablation |
FAR -- Multi-Cue, fixed fusion |
the visual and metadata channels mixed with a fixed weight of 0.5 instead of the learned gate |
jepa, frozen_encoder, encoder_mode=frozen |
FAR (frozen encoder) |
the retriever's visual encoder frozen at its contrastive pre-training weights |
l0 / l1 in run names |
-- |
retriever loss weight 0 / 1 (model.retriever_loss_weight) |
temporal |
Temporal |
the most-recent-frames baseline (TemporalRetrieval, K frames spaced interval seconds apart) |
worldmem |
WorldMem |
the FOV-overlap greedy retrieval baseline re-implemented from WorldMem |
longlive |
LongLive-RAG |
the content-query retrieval baseline (retriever trained without the action) |
agent_fa, rev_agent, A_nocue / B_rag |
dynamic-agent arms |
AI2-THOR-dyn arms whose generator also predicts the second agent's state; A = metadata scorer without the agent-state cue, B = with the cue |
| Name |
Meaning |
cdit |
Context-conditioned DiT (from Navigation World Models) |
cdit_b_2_mod_pose_fa (posefa) |
the main generator: full-attention CDiT with per-token Plücker-ray pose and trajectory-time conditioning; b = base size, 2 = patch size |
cdit_b_2_agent_fa |
pose_fa + the agent-state predictor (dynamic-agent corpus) |
ddim20, ps2 in run names |
DDIM with 20 sampling steps; "ps2" = the SoundSpaces pool/sampler setting used for all SoundSpaces arms |
| Name |
Corpus |
loopnav, loopnav_latent |
LoopNav (Minecraft loop navigation); _latent = pre-computed ViT-VAE latents |
ss, soundspaces_latent |
SoundSpaces v1: exploration with endpoint scans only |
ssv2, soundspaces_v2_latent |
SoundSpaces v2: exploration with periodic 360-degree scans (the headline SoundSpaces corpus) |
thor3, ai2thor_v3_latent |
AI2-THOR v3: object-interaction tours with a retrace (the headline AI2-THOR corpus) |
thor2, ai2thor_v2_latent |
the earlier AI2-THOR corpus (retriever pre-training data) |
thordyn, dyn, ai2thor_dyn_latent |
AI2-THOR-dyn: the door-corridor corpus with a second, moving agent |
_ctrl |
the static control sibling of each AI2-THOR v3 test episode (same stations, nothing moved) |
| Term |
Meaning |
| battery |
one evaluation sweep: every arm rolled out on every test clip, one .npz record per (arm, clip) under results/rollout_eval/<battery>/<arm>/ |
| arm |
one trained model (a memory variant) in a battery |
| replace-tail rollout |
autoregressive rollout in which the generated frame replaces only the tail slot of the context; the retrievable history stays real |
| context size / pool |
the number of frames the generator conditions on (generator.context_size) versus the number of past frames the memory may choose from (context_pool_size) |
| RAG step, EMDR² |
the retriever's training signal: the diffusion loss of the generator under each candidate retrieval set, turned into a posterior the retriever is distilled towards |
| stale frame, stale-margin success |
the exploration-phase view of the same viewpoint before an interaction; a step succeeds when the generation is closer to the ground truth than that stale view is |
chunk score, context_chunk_score |
the retriever's per-chunk relevance score, injected into the generator's context tokens |