memory_update was bound to `emitted_at(t, style=memory)`, which requires
the frame's exact timestamp to match a memory annotation. Memory rows are
placed at subtask-boundary timestamps and at 30 fps that's ~1% of frames,
so 99% of memory_update draws couldn't render and silently fell through
to _fallback_low_level_render — injecting task-conditioned low-level
training on ~30% of samples (subtask_mem.yaml).
Switch to `active_at`. At inference `MemoryUpdateFwd` is triggered on
`subtask_change` events, but the model only needs to learn the stateless
mapping (prior_memory, completed_subtask) -> current_memory. active_at
supervises this mapping on every frame inside a subtask interval, against
varied observations; the trigger lives outside the model. Net effect:
memory_update renders on ~87% of frames, the fallback leak drops from
~30% to ~4%, and memory CE gets a meaningful (not 0.3%) training share.
subtask_mem.yaml: rebalance to 0.30 / 0.55 / 0.15 so memory CE is
~13% effective and the freed weight goes to low_level_execution.
subtask_mem_vqa_speech.yaml: keep weights (memory_update=0.10 was
already balanced against the other text-CE branches).
Co-authored-by: Cursor <cursoragent@cursor.com>
When fine-tuning from pi05_base, reuse only the pretrained weights so pi052 still generates recipe text labels and FAST action labels.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep the PaliGemma LM head in float32 and initialize it from pretrained weights or token embeddings when loading pi05 checkpoints.
Co-authored-by: Cursor <cursoragent@cursor.com>
The parity check in debug_text_predictions was producing false ✗
DIVERGED reports. Root cause: I built the "inference" batch by
zero-masking the attention past the supervised span, but kept the
full 512-token padded sequence. select_message reads the prompt-end
hidden state via ``vlm_out[:, -1:]`` — the LAST position of the
prefix — which in a padded batch is a padding-token hidden state,
not the last prompt token. PaliGemma's prior on those padded
positions reliably argmaxes to <loc0879>, falsely flagging a
training/inference mismatch.
Fix: truncate both tokens AND mask to length == first_sup before
calling select_message, mirroring what the real runtime does
(``tokenizer(prompt)`` returns un-padded ids). Now the parity check
compares like-with-like.
The actual training argmax in the dump was sensible English
("' move the blue cube into the green bin'" at acc=6/9) — the head
is learning correctly. The "<loc>" salad was purely the harness
reading from the wrong position.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
At training time the policy is wrapped by Accelerator/DDP into a
.module attribute and custom methods are NOT proxied through the
wrapper, so ``hasattr(policy, "debug_text_predictions")`` was False
and the periodic dump was silently no-op'ing. Walk through .module
indirection to reach the raw PI052Policy that defines the method.
Also surface why the dump didn't fire (no method / empty supervised
positions / generation error) so users can see what's blocking it
instead of staring at silence.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extends the periodic LM-head dump (LEROBOT_DEBUG_PREDS_EVERY) to ALSO
run select_message autoregressively on the same prompt prefix and show:
prompt : '<bos>User: ... Assistant: '
target (ground truth) : ' close the gripper ...'
training argmax (teacher-fed) : ' close the gri lift ...' acc=12/15=80%
inference (autoregressive) : ' close the gripper around ...'
first-token parity : train=3387 (' close') vs infer=3387 (' close') ✓ MATCH
The first-token parity check is decisive: training-side argmax at the
prompt-end position and inference's first generated token both compute
``argmax(lm_head(h_last_prompt))`` on identical context, so they MUST
match. Any divergence signals a training↔inference bug (mask, dtype,
KI routing, embedding scale, etc.). Subsequent tokens can diverge
because training uses teacher forcing while inference free-runs.
debug_text_predictions now also returns an ``inference`` list keyed
by sample, each entry carrying ``first_sup_pos`` and ``decoded``.
Limited to 24 new tokens per sample to keep the dump fast.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds an opt-in diagnostic that, every N training steps, dumps 5 batch
samples plus the LM head's argmax prediction at every supervised
position alongside the label and a ✓/✗ marker — the cheapest signal
for "is text training actually learning what we expect, or collapsing
to a fixed token". Refills the recipe-sample dump budget on the same
cadence so the raw input shapes are also re-dumped.
Opt in via env var:
LEROBOT_DEBUG_PREDS_EVERY=1000 lerobot-train ...
PI052 implements ``debug_text_predictions`` (mirrors the text-loss
forward but returns argmax instead of CE); other policies are silently
skipped. The dump runs in eval() mode under no_grad, slicing the
current batch to N samples — no extra data fetch, no train-state
mutation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The trained model collapsed to spewing 40+ <loc> tokens for *every*
prompt — subtask, memory, anything — because VQA targets were supervised
to *start* with <loc>. With ~25% of all text samples beginning with a
<loc> token, the LM head learned "Assistant: → <loc>" as a strong
attractor; once one loc is emitted, autoregression chains the rest.
Flip the format so every text target — subtask, memory, speech, AND VQA
— starts with a regular word. The model still learns the <loc>
vocabulary for the spatial portion of the answer, but loc can no
longer be the first generation step out of a clean prompt.
Examples:
point : "green box <loc0162><loc0759>"
bbox : "cube <loc0082>…<loc0409>"
multi : "blue <locs> ; yellow <locs>"
The runtime parser (parse_loc_answer) strips loc tokens and uses the
remainder as label, so it's order-tolerant and works under either
format. Old loc-first checkpoints still parse cleanly at inference;
new training will use label-first.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
select_message's bf16 cast used next(paligemma.parameters()).dtype,
which lands on a fp32-kept param (norm / embedding) under
to_bfloat16_for_selected_params. Mask stayed fp32 while q/k/v were
bf16 → SDPA still raised "invalid dtype for bias". Read the dtype
from layers[0].self_attn.q_proj.weight instead — q_proj is always
cast with the rest, so its dtype matches what SDPA sees.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`_prepare_attention_masks_4d` always returns fp32 (the 0.0 / -inf
literals); with bf16 weights, HF PaliGemma's SDPA path raises
"invalid dtype for bias - should match query's dtype" and
select_message returns empty every step. Cast in both attention
sites: `_compute_layer_ki` (training, when both experts run) and
`select_message` (inference, VLM-only branch). Bf16 training +
bf16 inference now run end to end with no dtype mismatch.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
THE bug behind the <loc>-salad. PaliGemma's vocab reserves ids
[256000, 257023] for <locDDDD> detection / pointing tokens, but the
stock AutoTokenizer does NOT match them on raw text — it BPE-splits
<loc0162> into SEVEN pieces (<, loc, 0, 1, 6, 2, >). So a VQA target
like "<loc0162><loc0759> green box<eos>" tokenized to 16 pieces, not
5, and training the LM head supervised those generic BPE pieces
instead of one detection-vocab id. The piece logits got pumped up
across ~25% of supervised positions; at inference they dominated
every turn — even subtask prompts produced <loc>-salad followed by
the actual answer.
Register the 1024 <locDDDD> tokens via tokenizer.add_tokens once on
load, in every path the policy uses: PI052TextTokenizerStep (training
encode), _build_text_batch_pi052 (runtime encode), and
select_message's default tokenizer (runtime decode). Verified
empirically with the real PaliGemma tokenizer: VQA target now
tokenizes to 5 ids matching the loc-vocab range (256162, 256759, ...)
with correct offset_mapping.
This unlocks PaliGemma's actual detection prior; <loc>-salad cannot
recur because each <locDDDD> is a single class on the LM head, not a
character sequence the head accidentally learns to extend.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Confirmed empirically on the published dataset: VQA bbox/keypoint
coordinates are Qwen2.5-VL's 0–1000 normalized grounding output, NOT
pixels. Scanning 8207 samples showed x and y both spanning 0..1000
with ~30% of values exceeding the camera's pixel dimensions (which is
impossible if they were pixels).
_vqa_answer_to_loc was dividing by the observation image's H/W, so
e.g. point [742, 158] on a 640x480 wrist cam clamped x to <loc1023>
(the far-right edge) instead of mapping to <loc0760> (~74% across).
Fix: divide by 1000 — the actual Qwen scale. The conversion is now
camera-resolution-independent, so _camera_image_shapes and the
image_shapes plumbing through __call__ / _encode_messages /
_messages_vqa_to_loc are dropped. Tests updated to the new signature
and the 0–1000 round-trip.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
base.fit() rejected the data with "Vocab size 1024 is too small for
the range of tokens 9339": the FAST tokenizer was fit on raw
motor-unit actions, whose DCT-token range vastly exceeds the 1024
codebook.
Two problems, one fix. (1) Raw actions blow up the token range. (2) At
training time ActionTokenizerProcessorStep runs after the QUANTILES
NormalizerProcessorStep, so it encodes normalized actions — fitting on
raw actions mismatches that space. Replicate QUANTILES normalization
(per-dim [q01,q99] -> [-1,1], clipped) before base.fit() so the fit and
the training-time encode see the same distribution and the token range
fits the codebook.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
fit_fast_tokenizer collected action chunks via ds[i]["action"], which
builds a full training item — delta-timestamp expansion, video decode,
image transforms. A single video-decode failure threw, was swallowed
at debug level, and silently starved the fit of every chunk → "FAST
fit collected zero action chunks", falling back to the universal
tokenizer.
Read the ``action`` column straight from the HF dataset instead: it
carries no video, so it is immune to decode errors and far faster.
Also fail fast with a clear message when the dataset has no ``action``
feature or all episodes are shorter than chunk_size.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
_compute_layer_ki called modeling_gemma._gated_residual, but that
adaRMSNorm gated-residual helper is a lerobot helper in pi_gemma, not
part of HF transformers — so enabling knowledge_insulation crashed with
AttributeError on the first training step. Import _gated_residual from
pi_gemma, matching pi05's own layer code.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Spatial VQA answers (bbox / keypoint) were trained as pixel-coordinate
JSON, which fights PaliGemma's detection prior and leaks <loc>-token
salad at inference. Convert them to PaliGemma's native <locNNNN>
vocabulary instead so the LM head reuses that prior.
Training side (text_processor_pi052.py): a target turn whose content
parses as a bbox/keypoint answer is rewritten to <loc> text, using the
camera frame's native (H, W) from the observation and the preceding
image block. Non-spatial answers, subtask/memory targets and SmolVLA2
keep their JSON form — the dataset stays backbone-agnostic.
Runtime side (smolvla2/inference/vqa.py): parse_vqa_answer detects
<loc> answers (2 locs -> keypoint, 4 -> bbox), returning normalized
[0,1] coords with a normalized flag; draw_vqa_overlay denormalizes
against the chosen camera frame's pixel size.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Keep action-only samples trainable by rendering the task as a low-level user message when no recipe branch matches.
Co-authored-by: Cursor <cursoragent@cursor.com>
PI052TextTokenizerStep masked text_labels over the assistant turn's
*content only* — the trailing newline was excluded and no EOS token was
ever a supervised label. So the LM head was never given a stop signal:
at inference select_message decoded to max_new_tokens, producing the
runaway subtask paragraphs and the "}"}"}-style VQA tails.
_format_messages now appends the tokenizer's EOS to each supervised
target turn and extends that turn's span to cover it, so the EOS lands
in text_labels. _shifted_ce then trains "<last content token> -> EOS"
and the model learns to terminate; select_message stops on it.
Inference callers (the runtime's _build_text_batch_pi052) pass no
target_indices / eos_token, so no EOS is baked into the prompt — the
model generates it. Verified end-to-end with the PaliGemma tokenizer:
the supervised span is `<content><eos>` and the trailing newline stays
unsupervised.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The runtime's text path was hard-wired to SmolVLA2: _build_text_batch
read policy.config.vlm_model_name (which PI052Config doesn't have) and
built a SmolVLM2 chat-template prompt. PI052/PaliGemma is not
chat-pretrained and trains on a flat `User: ... \nAssistant: ...`
prompt, so the runtime crashed or fed an out-of-distribution prefix.
- _build_text_batch now dispatches on policy.config.type: smolvla2 ->
chat template (renamed _build_text_batch_chat); pi052 -> flat
role-prefixed text via PI052TextTokenizerStep's own _format_messages /
_strip_blocks / _flatten_say_tool_calls, so the inference prefix
matches PI052 training exactly.
- Add a lerobot-pi052-runtime entry point (alias of the same main; the
policy type is read from the checkpoint) so the command name isn't
misleading. argparse prog now defaults to the invoked command name.
PI052's select_message / predict_action_chunk already work with the
runtime; this was the one SmolVLA2-only coupling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The first-person memory narrative, task-rephrasing and initial-speech
prompt tweaks belong in the annotation pipeline itself. Applied to
feat/language-annotation-pipeline (#3471); reverting them here to the
merge-base so they drop out of this PR's diff. general_vqa.py keeps its
docstring fix since it references a recipe this PR introduces.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The deterministic-plan rewrite, single-frame VQA (K 3->1), dataset
version tagging, telegraphic-subtask prompt and shorter interjection
prompt belong in the annotation pipeline itself, not in the SmolVLA
training PR. They have been applied to feat/language-annotation-
pipeline (#3471). Reverting these six files here to the merge-base so
they drop out of this PR's diff; #3491 will inherit the canonical
versions when it next rebases on its base.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Convert PI052 prefix-only attention masks before calling PaliGemma so text-only batches and generation use the same mask shape as fused training.
Co-authored-by: Cursor <cursoragent@cursor.com>
Route full PI05/PI052 fine-tuning through PyTorch's fused AdamW path to avoid the single-tensor Adam denominator allocation near GPU memory limits.
Co-authored-by: Cursor <cursoragent@cursor.com>
Avoid the multi-tensor AdamW temporary that can OOM full PI05/PI052 fine-tuning near GPU memory limits.
Co-authored-by: Cursor <cursoragent@cursor.com>
Select only supervised text and FAST action-code positions before cross-entropy to avoid full-vocabulary loss tensors over padded sequences.
Co-authored-by: Cursor <cursoragent@cursor.com>
Use PI05Policy helpers for action padding and image preprocessing in PI052 fused losses instead of looking them up on the inner PI05Pytorch module.
Co-authored-by: Cursor <cursoragent@cursor.com>
Tokenize batched recipe outputs in PI052 so training batches with nested message lists do not crash before model forward.
Co-authored-by: Cursor <cursoragent@cursor.com>
Mask the FAST auxiliary loss to discrete action-code tokens so wrapper formatting tokens do not affect action co-training.
Co-authored-by: Cursor <cursoragent@cursor.com>
Module 3 anchored each VQA emission tick to K=3 consecutive frames
(~0.1s at 30fps). The VLM grounds the answer — bbox/keypoint
coordinates especially — against the first frame's image, so copying it
onto frames 2-3 smears a stale label over a moving scene.
Default K=1: a VQA pair lands on exactly its emission frame, no
temporal smear. VQA frames get sparser; the WeightedEpisodeAwareSampler
(vqa_target_fraction) is the knob to compensate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- hirobot.yaml -> subtasks_vqa.yaml
- hirobot_memory.yaml -> subtask_mem_vqa_speech.yaml
- pi05_hirobot.yaml -> deleted (stale: uses plan, top-camera names;
superseded by the two recipes above)
- smolvla2_hirobot.yaml -> deleted (was untracked stale junk)
Updated the smolvla2 / pi052 `recipe_path` config defaults, all
docstring / comment references, the annotation-pipeline + recipe docs,
and the three tests that loaded pi05_hirobot.yaml (repointed to the
renamed recipes; the low-level-branch and pipeline-render assertions
now accept a flow-only `low_level` stream as valid supervision, since
the new recipes' low_level_execution has no text-CE target).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
pi052 had the same text-CE collapse bug smolvla2 had — PaliGemma's
embed_prefix flags the language block att=0, so make_att_2d_masks makes
it fully bidirectional and the text cross-entropy degenerates into a
copy task. Ported the three model-specific fixes:
- _mark_target_span_causal: set att=1 on supervised target language
positions so the text-CE is genuine causal next-token prediction.
Applied in both _compute_all_losses_fused and _compute_text_and_fast_loss.
- flow_loss_weight 10.0 -> 5.0: the paper's a=10 swamps the LM head once
the flow-only low_level recipe fires often (matches SmolVLA2Config).
- _flatten_say_tool_calls in the text tokenizer: serialize `say` tool
calls into a <say>...</say> marker so the spoken reply is tokenized
and supervised (PaliGemma's flat prompt has no structured calls, so
they were dropped entirely).
select_message needed no change: pi052's prefix is [images, language]
with no trailing state token, so it already decodes from the last
language token.
Regression tests mirror the smolvla2 attention-masking + tool-call suite.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
VQA annotations are sparse, so VQA was badly underrepresented in training:
its effective share was weight x density, and blend draws that picked an
ask_vqa* sub-recipe for a non-VQA frame were wasted entirely.
Two pieces:
1. Recipe-side consumption (language_render.py): render_sample now routes
any frame that carries a VQA annotation to a matching ask_vqa* sub-recipe,
regardless of the weighted blend draw. No VQA annotation is wasted and no
draw lands on a non-renderable VQA recipe — VQA's recipe-side share now
equals the VQA-annotation density.
2. Dataset-side oversampling (WeightedEpisodeAwareSampler + vqa_target_fraction):
a new weighted, episode-aware sampler draws frames with replacement by
per-frame weight. When TrainPipelineConfig.vqa_target_fraction is set, the
train script scans language_events, weights VQA frames so they make up
~that fraction of the training stream, and uses the weighted sampler. This
is what actually lets VQA exceed its natural density. Default None keeps
uniform episode-aware sampling unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The VLM already sees every camera, so the operator never needs to name
one to ask a question. Move the camera prompt to after generation and
only fire it when the answer actually carries a bounding box / point
(whose pixel coordinates are camera-specific and need a target frame).
Non-spatial answers (count / attribute / spatial / plain text) now skip
the prompt entirely.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
handle_vqa_query filtered the observation down to the single chosen
camera before calling the VLM. But training feeds every camera: the
ask_vqa_* recipes' image blocks are stripped before tokenization and
the frames reach the model via OBS_IMAGES_*, where embed_prefix
consumes all config.image_features regardless of the per-camera recipe
tag. Filtering to one camera changed the image-token count in the
prefix (the dropped camera zero-padded with mask=0) — a prefix shape
the model never saw at training.
Now the full observation is passed to select_message; the chosen
camera is used only to pick which frame the bbox/point overlay is
drawn on.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Command arguments never needed quotes (`_strip_quotes` only strips a
matching pair if present) — `/question point to the yellow cube` works.
The hints wrongly implied `""` were required; all hints/help now show
`/action <task>` / `/question <text>`.
Also adds a reference line to the state panel showing the two
overlay-producing VQA prompt shapes:
/question point to the yellow cube -> point overlay
/question detect the blue cube -> bounding-box overlay
plus the same examples in /help.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replace the startup mode prompt + task picker with a single
command-driven prompt. The runtime now comes up immediately at the
command line in `paused` mode (robot idle) and the operator drives it:
/action "task" run the robot on a task (bare = resume, number = timed burst)
/pause stop the action loop — robot holds position
/question "..." pause and answer one VQA question (camera prompt + overlay)
/help / stop
- Removed _select_mode_interactively / _select_task_interactively /
_dataset_task_strings (the interactive pickers).
- mode value renamed "question" -> "paused"; --mode choices are now
action|paused (default paused).
- /question takes the question inline and runs it via _handle_slash_command
(pauses first, so the policy isn't used concurrently).
- The ENTER-to-start gate only fires when starting in action mode.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Lets the operator skip the interactive startup entirely and go straight
to the command line:
- New --mode {action,question} arg; when given, the startup mode prompt
is skipped.
- When --task is passed explicitly on the CLI, the startup task picker
is skipped (the dataset-bootstrap task still shows the picker so you
can override it).
Also adds a timed action burst: /action <seconds> runs the robot for N
seconds, then the autonomous loop auto-reverts to question mode and
clears the action queue. Plain /action stays unlimited.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New recipe alongside hirobot.yaml (kept as the lean baseline). Superset
that adds two text-supervised sub-recipes:
- memory_update: compress progress into a memory note.
- user_interjection_response: reply to a user interjection with a `say`
tool call only (no plan/subtask text). The SmolVLA2 chat tokenizer
flattens the call to a `<say>...</say>` marker the runtime parses back.
Plan is intentionally omitted; memory is the only persistent high-level
state. Weights: low_level 0.40, subtask 0.25, memory 0.10, interjection
0.10, vqa 0.075 x2.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a mode prompt at startup, shown before the task picker, so the
operator chooses action (run the robot) vs question (VQA only) up front
instead of having to discover /vlm mid-run.
Also rename the VQA mode from "vlm" to the clearer "question":
- state["mode"] value is now "action" | "question"
- the command is /question (/vlm and /vqa kept as aliases)
- panels, hints and help text updated to match
handle_vqa_query now reports via both push_log and direct stdout, so
VQA answers / overlay paths are visible in autonomous question mode
where the panel redraw is suspended.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The autonomous panel redraw cleared the screen every 0.5s, so the "> "
prompt and the one-shot command hint vanished — the operator could not
see what to type or what they were typing, making /vlm unreachable.
- Suspend the timer redraw entirely while in /vlm mode (the action loop
is paused, nothing changes in the background) so the VQA question and
camera prompt stay on a stable screen.
- Re-print the "> " prompt after each redraw so it is always visible.
- Show an always-on command hint in the panel (/vlm, /help, /action)
instead of relying on the startup line that scrolls away.
- Redraw immediately after a slash command so the mode flip is visible.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The picker was skipped whenever a task was already resolved — which is
always the case with --dataset.repo_id, since the dataset's canonical
task is auto-filled. The operator never got to choose. Now the picker
always runs on an interactive terminal: the resolved task is shown as
"(current)" and selected by an empty Enter, so the dataset-canonical
default still works while letting the operator pick another task or
type a custom one.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three additions to the SmolVLA2 interactive runtime:
1. Startup task picker — when no --task is given, the runtime lists the
dataset's task strings as a numbered menu (plus a custom-task option)
instead of silently waiting for the first stdin line.
2. Mode toggle — /action and /vlm slash commands flip a persistent run
mode. /vlm pauses the whole action loop (HighLevelSubtaskFwd,
LowLevelForward and DispatchAction gate on state["mode"]) and clears
the action queue so the robot holds position; /action resumes it.
The mode is shown in the state panel.
3. Interactive VQA — in /vlm mode a typed line is a VQA question. The
new inference/vqa.py module asks which camera to ground on, runs the
VLM on that single camera, and when the answer is a bbox/keypoint it
draws the overlay, saves a PNG to ./vqa_overlays/ and auto-opens it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The chat tokenizer passed assistant `tool_calls` straight to
`apply_chat_template`, which renders them as a structured JSON
`<tool_call>` block — so the LM head was trained to emit JSON. But the
inference parser `_split_plan_and_say` looks for a `<say>...</say>`
marker, which the model never saw in training, so the `say` tool never
fired at inference.
`_flatten_say_tool_calls` is the missing training-time serializer (the
one `_split_plan_and_say`'s docstring already assumed existed): it
rewrites a `say` tool call into a `<say>...</say>` marker inside the
content text before the chat template runs, so the template only
tokenizes plain text and the supervised target span trains the model to
emit exactly the marker the runtime parses back (Pi 0.5-style flat
tool-call serialization).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
SmolVLA's 1e-4 is safe only because it freezes the language head. SmolVLA2
unfreezes lm_head + the last text layer and fine-tunes the pretrained
SmolVLM2 language weights; 1e-4 is too aggressive there and destabilises
generation into degenerate repetition. Match pi05's 2.5e-5 peak LR.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The Pi 0.5 α=10 split assumed text is a rare auxiliary task. With the
flow-only `low_level` recipe (~40% of the blend) now rendering, the flow
term fires often and at 10x weight dominates the shared VLM backbone,
starving the text head into degenerate repetition decoding. A 5:1 split
keeps actions primary while leaving the language head enough gradient.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A recipe whose only supervision is the action-expert flow loss (e.g.
`low_level_execution`: `user(${subtask})` with `stream: low_level` and no
`target` turn) was rejected at render time by `_render_message_recipe` and
`_validate_rendered`, both of which required at least one target turn.
The result: every blend draw of the flow-only recipe rendered to `None`,
`predict_actions` was never set, `run_flow` never fired, and the action
expert received no flow loss — leaving it at random init. Both gates now
also accept a `low_level`-stream turn as valid supervision.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Regression coverage for the text-CE collapse bug fixed in 3cd348ff.
Pure-function tests over ``_mark_target_span_causal`` /
``_locate_lang_range`` / ``make_att_2d_masks`` — no model load, fast.
Pins:
* the target span flips to att=1, prompt/images stay att=0;
* target tokens attend causally among themselves (no peeking at
future targets) — genuine next-token prediction;
* targets still attend bidirectionally to images + the user prompt;
* the action-expert (state) token still attends to every target;
* a no-target subtask (low_level_execution user turn, labels all
-100) leaves the mask bidirectional;
* an explicit test documenting the bug: the raw embed_prefix mask
lets the first target token see the last — the copy-task collapse.
Skips cleanly when transformers isn't installed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
pepijn