Skip to main content

inspect-robots-dropbear

An Inspect Robots policy adapter for Dropbear-hosted DreamZero-YAM. Discovery and construction are offline; the first trial reset opens one lazy Dropbear connection, and later trials reuse that connection while starting fresh logical episodes.

Licensed under Apache 2.0. It supports Python 3.11 through 3.13 and requires the immutable dropbear[dreamzero]==0.1.0a9 SDK release.

Using it against a Dropbear-hosted model needs an API key and an entitlement for that model; the adapter itself is open.

Install and discover

Install the pinned release:

uv add "inspect-robots-dropbear @ git+https://github.com/Dreamscale-Labs/inspect-robots-dropbear.git@v0.1.6"

Confirm the expected Dropbear SDK is active before starting an evaluation:

python -c 'import dropbear; assert dropbear.__version__ == "0.1.0a9"'

Verify that the entry point is available without opening a cloud session:

inspect-robots list policies

The output must contain dropbear. Robocurve keeps its existing registered task and embodiment; do not replace or rename either. Change only the policy selection in the existing evaluation command:

--policy dropbear -P model=dreamzero-yam

The default is YAM's qualified async_latest mode. Use sampling=async_8 for the explicit compatibility/rollback path and sampling=upstream_eval only for the agreed open-loop dataset evaluation. The server keeps inference single-flight and latest-only. The SDK preserves two committed steps and applies its fixed absolute-target motion smoother only to the aligned async_latest suffix; the adapter exposes no custom scheduling, buffering, horizon, or smoothing knobs.

The first connection has a 1,800-second startup budget by default so DreamZero-YAM can finish loading and warmup. Set -P startup_timeout_s=<seconds> to another finite positive value when an approved target needs a different startup budget. This does not change timeout_s, the existing per-step action deadline, which remains 60 seconds by default.

Observation and simulator contract

The existing task and embodiment must provide all of the following on every policy step:

  • top_cam, left_cam, and right_cam uint8 images, each with a real monotonic capture time;
  • finite packed joint_pos state with shape (14,) in YAM left-arm, left-gripper, right-arm, right-gripper order; and
  • Inspect's integer extra["env_step"], starting at zero and advancing once per delivered action.

The adapter declares a 14-dimensional raw absolute-joint action at 30 Hz. It returns exactly one action per Inspect act() call while Dropbear owns DreamZero's managed action buffering. Simulator compatibility means matching those camera, state, action, clock, and rate contracts; it does not by itself establish physics parity, task success, or physical-robot safety.

Artifact ownership and joining

Inspect remains canonical for the EvalLog, aggregate scores, post-approval commanded-action JSONL, stored frames, Rerun recording, operator judgement, and trial termination/error state. The adapter adds one atomic diagnostics sidecar and records its relative path at TrialRecord.metadata["dropbear_telemetry"]:

dropbear/<run_id>/<sanitized-scene-id>-e<epoch>.jsonl

Schema-v2 sidecar rows contain serving identity, source control tick, source camera capture-to-execution age, timing, accurate maximum overlapping-target revision, chunk/merge disposition, and the same Inspect environment step. Join them to the EvalLog, action JSONL, or Rerun timeline using env_step; use join_key (<cache_generation>:<logical_action_index>) for Dropbear chunk diagnostics. Sidecars do not duplicate action vectors, images, credentials, authorization material, certificates, or endpoints.

Deterministic cleanup

Evaluation owners must call policy.close() in finally after the run, even when Inspect reports an error or cancellation. Then verify that the just-used session is gone:

dropbear sessions list

Do not stop unrelated sessions. The adapter also registers a bounded process-exit fallback, but it is not a substitute for explicit close.

Physical YAM boundary

This integration does not authorize an unattended physical run. For any physical YAM test, Robocurve owns and must supply its inspect-robots-yam package, validated limits, an attended operator gate, a working e-stop, and a rehearsed termination procedure. DreamScale's adapter calls neither the embodiment nor hardware directly, and compatibility or serving evidence must not be reported as physical safety or effectiveness evidence.

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

inspect_robots_dropbear-0.1.7.tar.gz (80.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

inspect_robots_dropbear-0.1.7-py3-none-any.whl (15.7 kB view details)

Uploaded Python 3

File details

Details for the file inspect_robots_dropbear-0.1.7.tar.gz.

File metadata

  • Download URL: inspect_robots_dropbear-0.1.7.tar.gz
  • Upload date:
  • Size: 80.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: uv/0.12.5 {"installer":{"name":"uv","version":"0.12.5","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for inspect_robots_dropbear-0.1.7.tar.gz
Algorithm Hash digest
SHA256 c9ffc0020f4db2849e1fd8acbc639b52c0c99bfa71423ce0c1b84ec345a6c740
MD5 7330671d937e954c604e65464dbf1da2
BLAKE2b-256 31dceff1fffa2b581470ce02bd2e6b4cd6828647446b135310abeaa995f1d61d

See more details on using hashes here.

File details

Details for the file inspect_robots_dropbear-0.1.7-py3-none-any.whl.

File metadata

  • Download URL: inspect_robots_dropbear-0.1.7-py3-none-any.whl
  • Upload date:
  • Size: 15.7 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: uv/0.12.5 {"installer":{"name":"uv","version":"0.12.5","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}

File hashes

Hashes for inspect_robots_dropbear-0.1.7-py3-none-any.whl
Algorithm Hash digest
SHA256 3c09e1e148bf1551454e66c26f0b93453af820df4fc9732e180e4edceb528dee
MD5 d7f23fe87610673d410ec74fcb409e33
BLAKE2b-256 7f26ab3023a2e903eb28a939c20361beace3f26aac5646fb16c99273f8c53b4f

See more details on using hashes here.

Release history Release notifications | RSS feed

0.1.17

2 files

0.1.16

2 files

0.1.15

2 files

0.1.14

2 files

0.1.13

2 files

0.1.12

2 files

0.1.11

2 files

0.1.10

2 files

0.1.9

2 files

0.1.8

2 files

This release

0.1.7 This release

2 files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page